Changelog

All notable changes to roam-code will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Perf + Pattern-1 stabilisation campaign (2026-05-21)

Performance

• intent 66s → 12s. _scan_doc_for_symbols replaced the O(lines × symbols) per-name re.search hot loop with one combined \bNAME\b alternation regex per line + cheap set-membership lookups. Output-identical (symbol_names is a set, so ref order is unchanged) — verified by construction + 17 passing tests.
• doc-staleness 93s → 19s. The per-file git blame walk was sequential; it now runs across a bounded ThreadPoolExecutor (min(8, cpu_count())). executor.map preserves input order, so the result list and its final stable sort are byte-identical to the serial version (verified A/B excluding the file-under-test's own self-blame).
• sbom 30s → 9s. sbom_reachability.py's filesystem scanners called Path.rglob('*') and discarded .roam/.git/node_modules post-hoc — descending into .roam's ~56K paths on every scan. A shared _walk_pruned helper now prunes skip-dirs during os.walk descent; output-identical (same skip set → same yielded file set).
• alerts — latent O(n³) bounded. _check_trends fed the full snapshot history into an O(n)-window × O(n²)-Mann-Kendall test. The fetch is now capped at the 200 most-recent snapshots (ROAM_ALERTS_SNAPSHOT_LIMIT env override) — a no-op on normal repos, bounded on long-history ones.
• clones — degenerate clone_detect bucketing fixed. node_count // max(node_count // 3, 5) collapsed every function ≥ 15 AST nodes into a single size band (~3 bands total → near-O(n²) pairing). Replaced with geometric base-2 banding int(log2(node_count)); base-2 plus the existing ±1 adjacency scan provably enumerates every pair within the 0.5 node-count ratio gate, so clone pairs / clusters / similarities are output-identical (only the sequential cluster_id label renumbers).
• smells refused-bequest — N+1 bulk-hoisted. _detect_refused_bequest issued 2 SQL SELECTs per inherits edge; both are now bulk pre-fetched once via batched_in. Output-identical (verified byte-for-byte over 4494 findings).
• smells / health parallel-hierarchy — O(S²) → recall-preserving inverted index. The detector Jaccard-compared every ordered pair of eligible superclasses; it now builds a token → superclasses inverted index and only compares pairs sharing ≥ 1 marker token. Recall-preserving (any pair with Jaccard > 0 necessarily shares a token), so output is identical.
• pr-risk — per-file git diff collapsed to one call. _get_file_stat spawned one git diff --numstat -- <path> subprocess per changed file; replaced with a single git diff --numstat over the same range parsed once into a {path: (added, removed)} dict. Output-identical.
• verify-imports — per-miss full-table-scan eliminated. The miss path fired a leading-wildcard path LIKE (un-indexable full scan of files) per unresolved import; replaced with an O(1) pre-built basename-set lookup that exactly replicates the SQLite LIKE semantics (ASCII case-insensitivity, _ single-char wildcard). Output-identical (31,886 import rows verified byte-identical); a latent-scalability win for large monorepos.

Fixed

• Pattern-1 — 17 commands omitted isError/status on their error envelopes. affected-tests, check-rules, context, coverage-gaps, file, fitness, grep, history-grep, ingest-trace, intent-check, invariants, plan, preflight, refs-text, relate, report, and reset emitted a JSON envelope on their usage-error / failure path without the machine-gate fields isError: true + a closed-enum status. All 17 now emit them (fitness + check-rules conditionally — only when a rule actually failed; a clean pass stays clean). New drift-guard tests/test_pattern1_envelope_shape.py parametrically asserts the shape across the 15 unconditional commands.

Post-v13.3 execution wave (2026-05-20)

Added

• B6 — adversarial × MCP-sampling "Dungeon Master". roam_adversarial gains an async compress_mode parameter (closed enum off / digest / defend, default off so the tool stays a pure superset with zero behavior change). digest summarizes the adversarial challenges via Context.sample; defend threads a defender-framing system prompt into the sampling call. compress_with_sampling now takes a system_prompt kwarg so defend-mode's framing actually reaches ctx.sample (previously dead code). An unknown enum value sets a loud summary.compress_mode_invalid: true sentinel (Pattern-1D) while preserving the verdict (LAW 6). Counts unchanged (core 57 / registered 227 / commands 241).
• B8 — spectral-forecast block in roam forecast. The forecast envelope now carries a one-shot spectral_forecast block (instability + decay-rate + decay-alert wording) computed from the current graph's spectral gap, plus a topology_decay_rate_definition sidecar. Option-B implementation: no DB column, no USER_VERSION bump; the decay projection reports insufficient_history until a persisted gap series lands. Findings are envelope-only (forecast is invocation-scoped).
• Pre-push cascade-prevention gate. New scripts/prepush_check.py (--fast / --full, per-step timing, copy-paste fix command on failure) bundles ruff (format + check) + the count/render drift scripts + an 11-test AST-scan structural-lint set into a ~43s local gate, plus a thin .githooks/pre-push shim that delegates to it. Catches the "new code trips a repo-wide drift guard the author never ran locally" class behind this session's fix-forward cascade. New tests tests/test_prepush_gate_wired.py pin the hook → script wiring and assert every FAST-set test file exists. Brings scripts/ under ruff discipline (cleared 5 latent lint errors).
• Bounded all-command smoke harness (dev tooling). New dev/roam_smoke.py runs every canonical CLI command once, argless, in --json mode — each in its own subprocess with stdin=DEVNULL (immediate EOF, no console-handle hang) and a hard per-command timeout, on a small thread pool so one slow/hung command can't stall the sweep. A finite list run once = no loop; subprocess isolation = one crash never blocks the next. It is a HANG/CRASH detector first (argless invocation is intentional — a clean usage/error envelope is healthy Pattern-1 behavior); the failure classes it hunts are HANG, CRASH, and BAD_JSON / EMPTY_STDOUT Pattern-1C envelope violations. Output is incremental JSONL + a human summary; safe to Ctrl-C. This sweep caught the 11-command BAD_JSON usage-path class fixed below.

Fixed

• W1280 — feature-envy false-positive cut (~91%). The feature-envy detector was a pure cross-FILE outbound-edge-ratio heuristic (>=4 edges, >50% of targets in another file); a 24-sample dogfood measured ~88% false-positive / 0% true-positive — mostly tests/ files and Click command / emit_ / build_ / collect_ / _section_ orchestrators that reference many modules by design. The predicate now (1) skips test-role files (canonical is_test_path) + orchestrator/assembler-named functions, and (2) requires the external refs to be concentrated on a single foreign file (true envy) rather than spread across many. Result on roam-code: feature-envy 2036 → 177 rows (~91% drop, zero test-file leakage, not zeroed — 177 genuine single-foreign-file cases remain); total smells 7794 → 5936. The finding message now names the dominant-foreign-file share so the claim is self-explaining. FEATURE_ENVY_DETECTOR_VERSION 1.0.0 → 1.1.0; composite SMELLS_DETECTOR_VERSION 1.4.0 → 1.5.0. +3 regression tests (orchestrator-named, spread-across-files, test-role-file all assert no fire) alongside the existing genuine-envy positive case.
• dangerous-eval RegExp.exec() / declaration-line false-positives. The code-injection detector matched <regex>.exec( and function exec( / def exec( declaration lines as eval-class sinks (~80% FP on JS/TS in dogfood). Added guards that skip dotted .exec( unless the receiver is a shell-exec module (child_process / cp) and skip declaration lines. dangerous-eval DETECTOR_VERSION → 1.1.0.
• hotspots py-eval-exec identical false-positive (family-closer). Same dotted-.exec( and def exec( declaration-line guards applied to the py-eval-exec security sink in cmd_hotspots.py.
• god-components flagged dataclass fields as CRITICAL. classify keyed only on file path and never on symbol kind, so high-fan-in prop / field symbols (e.g. EvidenceArtifact.path) landed in actionable-CRITICAL god components. Added a kind-aware guard so non-logic kinds are excluded/down-banded.
• vibe-check boilerplate-inflation rate exceeded 100%. _detect_boilerplate_inflation mixed per-occurrence and per-file counts, yielding rates like 163.7%. Now counts per-file so the rate stays in [0, 1].
• W564 CI cascade seal. Removed the inline _SEVERITY_ORDER table in adversarial_compress.py; it now delegates to the canonical severity_rank (src/roam/output/_severity.py), closing the severity-drift lint gate the B6 prototype tripped.
• Pattern-1C — 11 commands dumped plain text on their argless --json usage path. In --json mode, the no-argument / usage-guidance path of preflight, plan, affected-tests, ask, relate, context, file, history-grep, ingest-trace, report, and skill-generate emitted Click usage text, a bare VERDICT: line, or raw markdown — non-JSON stdout that a wrapper-bridge JSON parser chokes on (collapses to COMMAND_FAILED). Each now routes its argless path through json_envelope() with a standalone summary.verdict (LAW 6) + state + partial_success + an imperative hint, matching the canonical cmd_grep usage-path shape; text output is unchanged. New drift-guard tests/test_json_usage_path_envelope.py pins the SHAPE (stdout parses as an envelope with command + summary.verdict) across all 11 plus the two pre-existing positive controls (grep / invariants). Surfaced by the dev/roam_smoke.py BAD_JSON sweep.
• health — fabricated algebraic_connectivity: 0.0 when numpy+scipy absent (Pattern-2 honesty). roam health --json exported a hard-coded 0.0 Fiedler value when the eigensolver substrate is missing, indistinguishable from a legitimate 0.0 disconnected-graph reading. A new fiedler_failed flag now distinguishes "couldn't compute" from a real measurement: the JSON exports algebraic_connectivity: null + a companion algebraic_connectivity_available: false (plus the existing health_algebraic_connectivity_failed: warnings_out lineage marker), at both the summary and top-level sites; text mode prints n/a (requires numpy+scipy). MCP _SCHEMA_HEALTH updated at both sites to declare the new boolean. +2 regression tests in tests/test_w607_m_cmd_health_warnings_out_envelope.py (unavailable→null, available→verbatim value).
• W1287 — shotgun-surgery detector re-implemented (~1472 → ~27 rows, wrong-axis FP). The detector fired on graph_metrics.in_degree > 7 — pure INBOUND popularity, ~100% FP / 0 TP in an 18-sample dogfood (the top hits were the codebase's best-factored shared symbols: conftest fixtures, open_db / json_envelope / to_json helpers, dataclass fields; 69% in tests/). High inbound reference count is good factoring, the opposite of the smell. Re-keyed onto Fowler's actual axis: the count of DISTINCT NON-TEST CALLER FILES referencing the symbol (file-SCATTER — how many separate files a change ripples across). Deliberately conservative behind a high _SHOTGUN_MIN_CALLER_FILES = 12 threshold + the W1280-style test-role / @property / dataclass-field / trivial-accessor exclusions, so a well-factored repo reports ~zero rows by design. SHOTGUN_SURGERY_DETECTOR_VERSION 1.0.0 → 1.1.0; composite SMELLS_DETECTOR_VERSION 1.5.0 → 1.6.0. Tests in test_smells.py rewritten: genuine file-scatter fires, concentrated-popularity (50 refs from 1 file) does NOT, test-role target does NOT, below-threshold does NOT.
• W335/W342 — impact's reported blast-radius count now agrees with preflight (Pattern-3a + cap disclosure). roam impact's default --max-callers 100 capped both the displayed dependents AND the reported COUNT, so affected_symbols silently understated the true radius and contradicted preflight's gate for the same symbol. The reported total is now the honest uncapped reach — computed by the IDENTICAL nx.descendants-over-reverse-graph computation cmd_preflight._check_blast_radius uses — while only the listed dependents stay capped for response size. A new BLAST_RADIUS_AFFECTED_TOTAL sidecar (in metric_definitions.py, surfaced as affected_metric_definition) names the shared computation so the parity is provable. The display cap is now disclosed LOUDLY (cap_applied + displayed + total + a verdict that reads listing 100 of N affected symbols … raise --max-callers to list more) instead of silently truncating the count (Pattern-1 variant-D lineage). Risk classification + reach-pct now run on the honest total too. New test_impact_total_agrees_with_preflight_over_cap (120-caller fixture) asserts impact↔preflight parity; the existing cap/depth tests split into displayed vs total.
• W837 — forecast + coverage-gaps emitted Pattern-2 success envelopes on an empty corpus. roam forecast on a freshly-indexed symbol-less repo produced a misleading "spectral failure band" verdict from a degenerate 2-node file graph (gap 0.0 read as is_failed) — a clean run when there was nothing to forecast. It now probes the symbol count directly (the indexer writes a snapshot row even for an empty corpus, so snapshot count is not a reliable empty signal), and on zero symbols replaces the verdict with an explicit no data to forecast — corpus empty line + state: no_data + partial_success: true, suppressing the spurious spectral clause. A repo WITH symbols but <3 snapshots stays honest (insufficient snapshot history, no partial flag). roam coverage-gaps had verdict-less envelopes on its no-gates / no-entry-points branches (summary.error only, no verdict → None to a LAW-6 consumer); both now carry a standalone verdict naming the absent gates/entry-points + partial_success: true + a closed-enum state (no_gates / no_entries) + agent_contract.facts. New tests tests/test_w837_forecast_empty_corpus.py + tests/test_w837_coverage_gaps_empty_corpus.py.
• understand ↔ health verdict-label divergence (Pattern-3a / LAW-6). roam understand labelled a 75/100 health score "healthy" via an inline >=70 cutoff while roam health calls 75 "Fair" (its band reserves "Healthy" for >=80) — one score, two contradictory verdicts across commands. New canonical src/roam/quality/health_band.py owns the single band table (>=80 Healthy / >=60 Fair / >=40 Needs attention / <40 Unhealthy, mirroring cmd_health._compose_verdict); understand now routes its score→label through health_band() and stamps health_band + a health_band_definition Pattern-3a sidecar on the envelope. quality/__init__.py documents the new shared metric alongside cycles / god_components / public_symbols. New tests/test_understand_health_band_parity.py.
• fan — test-role symbols/files crowded out the headline ranking (test/prod split). roam fan's #1 fan-in on roam-code was the invoke_cli conftest fixture (2438 refs) — pure test noise burying real production coupling. Mirroring cmd_uses' production/test scope split (both now classify each subject via the canonical is_test_file helper), fan drops test-role rows from the headline ranking by default and annotates every shown item with a scope field. A new --include-tests flag opts them back in. The drop is disclosed loudly — summary test_split / production_items / test_items / test_filtered, a text-mode NOTE: line, and a distinct all_filtered_tests empty-state verdict — never silent (Pattern-1-D / Pattern-2 lineage). File-mode oversamples 5× so the headline still fills after the filter. New tests/test_fan_test_prod_split.py.

[13.3] — 2026-05-19

MCP runtime security wave (2026-05-18)

• MCP-P0.1 — egress redaction at the wrapper boundary. redact_secrets_in_string + recursive redact_secrets_in_value walker now run at the _wrap_with_receipt boundary in src/roam/mcp_server.py, so every tool response is scrubbed before it crosses the MCP transport. Closed-enum redactions=("secret",) flag on the envelope; per-pattern detail under _meta.extra.redaction_details. New tests: tests/test_w_mcp_redact_egress.py (5 tests). Closes the egress half of the "Poison Everywhere" prompt-injection class inside the server, complementing the gateway-side controls described in dev/MCP-SECURITY-POSTURE.md.
• MCP-P0.2 — 4-mode policy enforcement at the MCP boundary. New helpers _evaluate_mcp_mode_policy, _resolve_required_mode_for_tool, and _build_mode_blocked_envelope in src/roam/mcp_server.py gate destructive MCP tool calls against the active read_only / safe_edit / migration / autonomous_pr mode. policy_decision is a closed enum: allow / deny / not_evaluated. New tests: tests/test_w_mcp_mode_enforcement.py (7 tests). The MCP layer now matches the gate posture the CLI has carried since the mode substrate landed.
• MCP-P0.3 — HMAC-link MCP receipts into the signed event ledger. Each McpDecisionReceipt sha256 is appended to a signed runs/ event so an offline roam runs verify extends its 4-state verdict with a receipt_integrity closed enum: ok / missing / tampered / not_linked. New verify_chain_with_receipts() in src/roam/runs/signing.py; hash-stable on pre-P0.3 chains so existing ledgers re-verify byte-identically. New tests: tests/test_w_mcp_receipt_hmac_link.py (9 tests). The receipt stream is now part of the same tamper-evident chain as the run events that produced it.
• MCP-P2.1 — dev/MCP-SECURITY-POSTURE.md published for gateway integrators (Interlock / Lasso / Portkey). 404 lines covering the inside-server vs gateway split, the 4-layer ownership table (redaction / mode-policy / receipt-integrity / transport-policy), and a cross-link to public Discussion #37. Frames what the server owns and what value the gateway layer adds on top, so integrators do not duplicate controls.
• MCP-P2.2 — McpDecisionReceipt JSON Schema export. New src/roam/evidence/mcp_receipt_schema.py emits a Draft 2020-12 schema with additionalProperties: false and bidirectional dataclass↔schema parity. New scripts/export_mcp_receipt_schema.py writes the schema to disk for downstream consumers. Closed-enum drift-guards on _POLICY_DECISIONS and REDACTION_REASONS keep the schema and vocabulary in lockstep. New tests: tests/test_mcp_receipt_json_schema.py (11 tests). External tooling can now validate receipts without depending on the Python dataclass.
• W420 wheel-layout helper consolidation. _package_file() in surface_counts now resolves package files via importlib.resources.files("roam") instead of walking parent/src/roam/, fixing the installed-wheel layout for every surface-counts caller (not just the two commands sealed earlier today). New regression test tests/test_surface_counts_wheel_layout.py (9 tests) pins the helper end-to-end so the next contributor cannot reintroduce the source-tree assumption.
• W462 landing-page tool-count drift. index.html / press.html / llms.txt carried a stale 224 MCP-tool count; all three now read the canonical 227 (matching roam surface --json and the in-repo headline). Drift-guard test already in place from earlier waves catches re-introductions.
• W641 pr-risk verdict suffix. BLOCK and REVIEW verdicts now carry an explicit risk_level <tier> suffix (e.g. BLOCK (risk_level critical)) so downstream regex consumers can split on the tier without re-parsing the envelope.
• Phase 4 landing-page polish. Hero on index.html collapsed; SEO keywords surfaced higher in the document; 5 cross-links closed orphan pages; architecture.html lists the 8 evidence questions as an ordered list with 6 deep-link anchors. Bug fix: removed a stray continue from the roam mcp-setup client-list loop where the client was never in _CONFIGS to begin with.

Added — 2026-05-18 session

• Three new MCP wrappers — roam_boundary, roam_test_hermeticity, and roam_compatibility are now callable as MCP tools. Catch public-by-accident exports + wrong-direction imports, non-hermetic AI-generated test patterns (network / time / random / fs / env / subprocess), and outbound surface regressions vs a baseline — all from your agent without dropping to the CLI.
• MCP output sanitization at the server boundary — every MCP tool response now runs through a secret-redaction pass on egress (_redact_result_for_egress in mcp_server.py). Closes the "Poison Everywhere" output-side prompt-injection class at the server, not just at the gateway. Closed-enum redactions=("secret",) flag on the envelope; per-pattern detail surfaced under _meta.extra.redaction_details (5 new tests, 213 broader MCP tests pass).
• Wheel-layout regression test — tests/test_surface_counts_wheel_layout.py (160 LOC, 9 functions) pins the surface-count helpers to resolve package files via importlib.resources rather than walking parent/src/roam/. Prevents the W420 follow-on regression that broke roam --json surface + --json capabilities in installed-wheel layout (caught by post-push wheel smoke).
• New wheel-smoke CI job + extended bundling drift-guards — wheel-bundling drift-guards (W554/W570/W610) now cover templates/audit-report/, taint_rules/, and languages.extractors/ so a missing-from-wheel asset trips CI on the PR, not in the field.
• Pattern-1 family A/B/C/D coverage — ~30 new test files spanning Pattern-1A cold-start guard, Pattern-1B wrapper-bridge JSON pass-through, Pattern-1C empty-stdout envelopes, and Pattern-1D degraded-resolution disclosure; plus W420 invariants, W543 edge-kind canonicals, W744 status-split parity, severity/confidence parity, and lease/permit warnings_out propagation.
• W744 SARIF + policy status split — the Suppression dataclass is now a discriminated union (SARIF suppression vs policy suppression); warnings_out is wired at 4 production callsites that previously dropped it on the floor.
• MCP coverage: 227 tools registered total — 3 net-new wrappers this session, all thin DISPATCH shims following the canonical wrapper pattern.

Fixed — 2026-05-18 session

• roam --json surface + roam --json capabilities in installed-wheel layout (SHIPPING bug) — the W420 cascade fix had migrated runtime _COMMANDS reads to the surface_counts.cli_commands() AST helper for plugin invariance, but the helper resolved package files via parent/src/roam/cli.py — which doesn't exist in site-packages/roam/. Both commands now use _package_file() (importlib.resources-based) so the wheel-installed and source-checkout layouts both work. Caught by post-push wheel smoke before any user hit it.
• W420 cascade closure — surface, capabilities, compatibility, and doctor all now read _COMMANDS via the AST helper, so the published command_count headline is plugin-invariant across all four surfaces.
• README polish, round 2 — v12 / v11 release history moved out of the README into the changelog where it belongs; install-section duplication eliminated; the decorative "Works With" bar dropped (nothing it said wasn't already in the docs); the "Why use Roam" section trimmed from 6 bullets to 3 highest-signal ones; new worked roam preflight handleSave output example; em-dash typographic pass for consistency.
• CI fix-forward — ruff format on surface_counts.py; loosened a BLOCK verdict assertion to .startswith("BLOCK") so the W641 risk-level suffix doesn't trip the check.

Changed — 2026-05-18 session

• Detector cohort 26 → 28 — boundary and test-hermeticity are now persisting detectors. roam findings list --detector boundary and --detector test-hermeticity work end-to-end (subject-kind: file, symbol; closed-enum kinds documented in CLAUDE.md cohort section).
• Landing-page polish across 9 pages — index.html hero collapsed and SEO keywords surfaced; architecture.html now lists the 8 evidence questions as an ordered list with 6 deep-link anchors; pricing.html + governance.html "maps to, not certified" wording fix; audit.html + trust.html primary CTA flipped and SOC 2 / ISO honest dates labeled consistently; mcp-usage.html + command-reference.html counts re-synced; getting-started.html roam mcp-setup client list cleaned up (stray continue keyword removed).
• llms-install.md refreshed — canonical 11-step agent loop now matches the CLAUDE.md surface; count headlines re-synced.
• Public Discussion #37 reply posted — answered the "Interlock MCP-security gateway" question on GitHub Discussions; framed inside-server redaction (what landed today) vs gateway-side policy enforcement (where products like Interlock / Lasso / Portkey live).

In flight today (uncommitted; will ride the next squash)

• 4-mode policy enforcement at the MCP boundary — wiring read_only / safe_edit / migration / autonomous_pr modes through mcp_server.py so the MCP layer actually gates destructive tool calls (today the gates live in the CLI layer; the MCP wrapper currently bypasses them).
• Public MCP security posture doc — dev/MCP-SECURITY-POSTURE.md for gateway integrators (Interlock, Lasso, Portkey) describing the inside-server controls and where the gateway layer adds value on top.

Added

• W420 — test(invariant): NEW tests/test_w420_surface_count_plugin_invariant.py (85 LOC) pins the command_count headline equal across plugin-loading triggers. Plugin discovery no longer changes the published headline.
• Phantom-annotation CI lint — test(changelog): NEW tests/test_changelog_phantoms.py (237 LOC, 6 tests). Mechanically blocks future dev/*.md phantom references in CHANGELOG; 3-line annotation radius.
• Phantom-CLI doc gate — test(docs): NEW tests/test_doc_no_phantom_cli.py (~215 LOC). Scans 9 phantom-CLI patterns across 3 tracked agent-visible doc surfaces (AGENTS.md / README.md / CONTRIBUTING.md); mutation-test validated.
• W391 — feat(ci): GitHub Actions roam-sarif + CodeQL co-deploy sample at src/roam/templates/ci/roam-sarif-with-codeql.yml (77 LOC). Discovery wired through cmd_doctor._github_template_registry() + cmd_ci_setup docstring cross-link; reachability 10/10.

Fixed

• W420 — fix(surface): cmd_surface.py command_count no longer changes under plugin loading. Migrated from runtime _COMMANDS to AST helper roam.surface_counts.cli_commands(). Published count headlines unchanged.
• agents_md generator phantom — fix(agents-md): src/roam/agents_md/generator.py:962 emitted phantom roam rules check --strict (rules is flat, has no check subcommand). Corrected to canonical roam rules --ci. Protects downstream-user scaffolds.
• README findings-registry gap — fix(docs): dark-matter aggregator missing from README's findings-registry list (architecture.html had 7 names; README had 6). Re-aligned.

Changed

• Detector cohort 26 → 28 persisters — cmd_boundary + cmd_test_hermeticity reclassified as legitimate persisters; 5 doc surfaces updated (CLAUDE.md cohort + rule + README + architecture.html + evidence-checklist.md).
• Substrate package count 11 → 12 — CLAUDE.md + AGENTS.md "11 substrate packages" corrected to "12" matching the 12-entry body (atomic_io / agents_md / constitution / db.findings / laws / leases / memory / modes / policy / quality / runs / world_model).
• Rule-pack uniformity — 7 mainstream + NOTE-tier packs (python / typescript / kotlin / go / java / rust / swift) share imperative-voice + WHY-block discipline. README rule-count table corrected: rust 12 → 30, swift row backfilled, 6 severity-tier miscounts fixed.
• Phantom-memo annotation campaign closed — 34 phantom dev/*.md references annotated. Combined with the new CI lint, the failure class is sealed.

Closed

• W847 + W759-W761 — cmd_preflight + cmd_attest + cmd_risk + cmd_invariants + cmd_bus_factor + cmd_complexity UPPER-case audits all internal-vocab (W762 drift-guard pins discipline; 54 hits / 0 envelope-slot bugs).

Added — W1103-arc + W489-family-closed + capability-invariants + structured_unknown_filter-FULLY-CLOSED + symmetric-emission-COMPLETE batch (post-CONSOLIDATE-21, 2026-05-17 /loop iteration N+22)

> ~17 SHIPPED + 3 BAIL/SHIPPED + 1 RESEARCH MEMO + 2 REAL BUGS > fixed across 7 themes: regex CLI toggle (W421), taint qualified_only > lint family FULLY CLOSED (W489-A + W489-A-followup), capability-axis > invariant lint + 2 REAL BUGS fixed (W365 family), structured_unknown_filter > family FULLY CLOSED (W1083-followup-3 multi-value sibling), > symmetric-emission family COMPLETE (W1100 + W1101 + W1102), 2 > test-rot diagnoses (W844-drive-by-2 + W1084), and pruning + W1117 > placeholder family FULLY CLOSED (W507 + W1117-followup-4). Plus 5 > stale-BACKLOG hits doc-pinned (W844 + W1007 + W1008 + W851 + W414b) > — discipline rule re-affirmed at operational cadence.

• W421 — feat(cli): -E/--regexp flag on roam history-grep + roam refs-text; regex toggle threaded to underlying matcher. +25 LOC; 53 tests pass.
• W489-A + W489-A-followup — feat(taint): qualified_only lint wired into envelope on cmd_taint (Option A: catch_warnings capture) + helper hoisted to shared src/roam/security/taint_rules_lint.py + wired on cmd_cga. 62 + 107 tests pass. W489 family fully closed.
• W365 family — test(mcp-lint): _TOOL_METADATA ↔ ToolAnnotations parity lint + 3rd-surface capability-registry cross-check (2854 pass) + 3 logical-entailment invariants (destructive→NOT ai_safe; deprecated↔maturity; task_required→mcp_expose; 46 pass). Entailment surface exhausted.
• W1083-followup-3 — feat(structured_unknown_filter): multi-value structured_unknown_filter_many sibling + cmd_math + cmd_smells migration; +302 helper / +113-53 cmd_math / +193-42 cmd_smells / +336 tests; 366 broader pass. structured_unknown_filter family FULLY CLOSED (single + multi + CLI dispatcher).
• W1083-RESEARCH — docs(research): multi-value helper design memo at dev/W1083-RESEARCH-multi-value-2026-05-17.md.

<!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->

• W1117-followup-4 — feat(help): final 2 placeholder normalizations on cmd_clones; W1117 family fully closed — ~32 normalizations across 5-wave arc.

Fixed — W1103-arc batch

• W365-followup — fix(capability): roam_reset + roam_clean destructive flag corrected at the capability-decorator layer (2 REAL BUGS surfaced by W365 parity lint). +9/-10 LOC + 85 LOC test; 42 pass.
• W1100 — fix(envelope): partial_success: true on malformed agent_contract:[] (sibling to the CONSOLIDATE-21 schema_violations[] envelope-root ship). +30 LOC; 6 new + 142 broader pass.
• W1101 + W1102 — fix(envelope): symmetric list_counts: {} + preserved_list_truncations always-emit on zero-cardinality paths. +14 effective LOC; 369 + 381 pass. Symmetric-emission family complete across the 3 envelope-root array slots.
• W507 — fix(enum): prune dead 'self-hosted' enum value (0 consumers). 91 tests pass.
• W1084 — fix(test): refresh test_test_scaffold_unknown_symbol_passes_through for W1278a Pattern-2c migration; 39/39 pass sequential + parallel. W-number-collision target separate from the W1084 cmd_ai_readiness / cmd_fitness arc at CONSOLIDATE-20.
• W844-drive-by-2 — fix(docs): 3 stale README-headline references swept across tests + docs. 186 tests pass.

Changed — W1103-arc batch

• structured_unknown_filter family FULLY CLOSED — the helper-substrate arc that opened at CONSOLIDATE-18 (W1068-W1083 Phase 2) and propagated through CONSOLIDATE-21 (W1083-followup Phase 3) closes cleanly at CONSOLIDATE-22 with the W1083-followup-3 multi-value sibling. Single + multi + CLI dispatcher all aligned on the same did_you_mean: [...] envelope shape.
• Symmetric-emission family COMPLETE — the 3 envelope-root array slots that previously omitted on zero-cardinality paths (list_counts + preserved_list_truncations + schema_violations) now always-emit {} / [] for envelope-shape symmetry. Pairs with the W1101 CONSOLIDATE-21 ship — this batch verifies the symmetry across the family.
• W489 qualified_only family FULLY CLOSED — the taint qualified_only lint that opened at W489 propagates end-to-end: lint wired into envelope on cmd_taint (Option A catch_warnings); helper hoisted to shared src/roam/security/taint_rules_lint.py; cmd_cga wired alongside.
• W1117 placeholder family FULLY CLOSED — the [VALUE] → <value> normalization sweep that opened at CONSOLIDATE-21 closes the final 2 sites on cmd_clones; ~32 normalizations across the 5-wave arc.
• Capability-axis entailment surface EXHAUSTED — the 3 logical-entailment invariants (W365-followup-2) close the capability-decorator surface; 2 REAL BUGS fixed at the surface (W365-followup).
• BACKLOG-drift discipline re-affirmed — 5 stale-pending hits this session (W844 + W1007 + W1008 + W851 + W414b); all 5 doc-pinned SHIPPED-PRE-CONSOLIDATE-22 with retro note. Discipline rule from CONSOLIDATE-21 holds: verify code state before dispatching from a [pending] flag alone.

Added — W1067 → W1102 arc batch (post-CONSOLIDATE-20, 2026-05-17 /loop iteration N+21)

> ~30 completions consolidating the W1067 → W1102 wave-arc. Seven > themes carry the batch: Pattern-1D helper Phase 2/3 propagation > (7 callsites), W1142 cap-hit disclosure family closure (7 commands), > Pattern 3a severity widening on cmd_smells + cmd_adversarial (family > TERMINAL), W1117 placeholder normalization sweep (22 commands), > symmetric envelope emission (W1100 + W1101), W350 OSCAL > authority_refs projection (closes evidence-question Q2 coverage), > and permit-vs-lease asymmetry documented in CLAUDE.md (W1071).

• W1142-followup-A/-B — feat(disclosure): cap-hit on 7 list-truncating commands (cmd_clones + cmd_debt + cmd_recommend + cmd_test_impact + cmd_supply_chain + cmd_agent_score + cmd_runs). Canonical em-dash text uniform across all 7. cmd_search_semantic BAILED.
• W1068-W1083 + W1083-followup — feat(filter): structured "did you mean X?" suggestions across 7 unknown-name handlers (5 adopt structured_unknown_filter + 2 adopt to_summary_payload).
• W350 — feat(evidence): OSCAL Assessment Results projects authority_refs[] as EXAMINE observations; closes evidence-question Q2 coverage on the OSCAL projection axis.
• W1005 + W1005-followup-B — feat(severity): cmd_smells + cmd_adversarial accept the W547 7-token canonical severity vocabulary. Pattern 3a severity family TERMINAL.
• W1117 + W1117-followup-2/-3 — feat(help): 22 commands normalize [VALUE] → <value> placeholder convention end-to-end.

Fixed — W1067 → W1102 arc batch

• W1100 + W1101 — fix(envelope): symmetric list_counts: {} emission on zero-truncation paths + schema_violations[] envelope-root surfacing on malformed agent_contract:[].
• W844-drive-by — fix(docs): README hero pytest tests/test_basic.py example rotted by W405 shallow-git default; refreshed.
• W844-drive-by-2 — fix(docs): 3 stale headline references swept across README + landing-page + docs.
• W414d — bail(probes): git_repo + python_project module-scope BAIL-BOTH; both probes are structurally inapplicable at module scope (captured with rationale).

Changed — W1067 → W1102 arc batch

• Pattern 3a severity family TERMINAL — the W1005 arc that opened at CONSOLIDATE-18 with cmd_smells primary + cmd_llm_smells followup-A now sweeps cmd_adversarial as the third (and final) high-signal site. All three commands accept the W547 7-token canonical severity vocabulary (info / low / medium / high / critical / blocker / unknown).
• CLAUDE.md "Permit-vs-lease expiry-filtering asymmetry (W1067)" sub-section — codifies why permits load expired entries (audit-completeness semantic — proof that authority was exercised at the time the bundle was emitted) and leases filter them (live conflict-resolution semantic — an expired lease no longer holds a subject).
• BACKLOG-drift discipline codified — W1007 / W1008 / W844 / W1100 surfaced as shipped-but-pending instances; discipline rule going forward: verify code state before dispatching from a [pending] flag alone.

Added — W1086-arc + Wave-B-TERMINAL + W478 + Pattern-1A-family batch (post-CONSOLIDATE-19, 2026-05-17 /loop iteration N+20)

> ~20 completions since CONSOLIDATE-19 (Section 65). SARIF > dashboard family TERMINAL milestone — the W1062 + W1062-followup > trio + W1062-followup-2/3/4 fan-out + W1087 lint substitute arc > that kicked off at CONSOLIDATE-18 closes cleanly across 12 wired > emitters + 39 catalogued emitters end-to-end. Four themes carry > the batch: (a) SARIF dashboard family TERMINAL at 12 wired > emitters + W1087 lint substitute (W1062-followup-3 + > W1062-followup-4 wire 6 more emitters; W1087 catalogues 13 WIRED > + 26 EXEMPT = 39 emitters end-to-end). (b) MCP outputSchema > 13-tool Wave B TERMINAL carry-forward + Wave C1 implementation > kickoff — Wave C1 lands the first compat-profile env-vars > (ROAM_MCP_COMPAT_STRIP_OUTPUT_SCHEMA + ROAM_MCP_COMPAT_STRICT) > plus a sidecar hoist drive-by; pairs with the new > MCP-COMPAT-PROFILE-ROADMAP planning memo. (c) Pattern-2 + > Pattern-1A empty-state arc closure — 8 detectors + 2 hard-cap > commands sealed (W805-followup-bundle + W1085 + W1086 + W1084). > (d) 3 research memos drafted across the arc — MCP-COMPAT- > PROFILE-ROADMAP new this session, MCP-OUTPUTSCHEMA-EVOLUTION + > DETECTOR-FP-RATE-METHODOLOGY carry-forward. Plus 10 stand-alone > polish items (W365 + W459 + W478 + W844 + W847 + W759 + W986 + > W462 + W1088 + W1038) + 1 BAIL (W851). Parallel in stature to > the Wave B TERMINAL milestone that CONSOLIDATE-19 carried.

• W1062-followup-3 — feat(sarif): wire _derive_finding_tags() on clones_to_sarif + smells_to_sarif + over_fetch_to_sarif (11 tests pass). pr_risk_to_sarif found N/A (deliberate W1147/W1148 omission).
• W1062-followup-4 — feat(sarif): wire _derive_finding_tags() on n1_to_sarif + missing_index_to_sarif + orphan_imports_to_sarif (12 tests pass; 12 emitters wired end-to-end). W1087 captured as substitute-rather-than-wire ship for the long tail.
• W1087 — feat(sarif-lint): NEW tests/test_sarif_tag_coverage.py (6 tests pass). Two-part PIN + ALLOWLIST drift guard; 13 WIRED + 26 EXEMPT = 39 emitters catalogued. SARIF dashboard family TERMINAL.
• Wave-C1 — feat(mcp): MCP compat env-vars ROAM_MCP_COMPAT_STRIP_OUTPUT_SCHEMA + ROAM_MCP_COMPAT_STRICT (7 focused + 188 broader tests pass). Drive-by sidecar hoist — audit-metadata _meta block had escaped the fastmcp gate; hoisted to canonical wrapper path.
• MCP-COMPAT-PROFILE-ROADMAP — docs(research): Wave-C planning memo at dev/MCP-COMPAT-PROFILE-ROADMAP-2026-05-17.md (compat-profile-emit + roam mcp doctor probe surface for client-side capability negotiation). <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• W805-followup-bundle — feat(detectors): Pattern-2 empty-state migration across 5 detectors (cmd_vibe_check + cmd_fingerprint + cmd_fan + cmd_dark_matter + cmd_conventions); 25 tests pass. Pattern-2 detector empty-state coverage now 8/8 effective sites.
• W1085 — feat(fingerprint): Pattern-1A hard-cap disclosure on cmd_fingerprint empty-state path (19 tests pass; W-number-collision target separate from the W1085 cmd_fitness arc).
• W1086 — feat(cut): Pattern-1A hard-cap disclosure on cmd_cut empty-state path (4 tests pass; mirror to W1085 fingerprint).
• W365 — test(mcp-lint): NEW CI lint cross-check _TOOL_METADATA ↔ ToolAnnotations (10 tests pass). Finding: ToolAnnotations FULLY wired today — W363 materially less critical than the -18 audit feared.
• W986 — docs(claude-md): codify "First hypothesis" test-failure-triage discipline rule (W978 + W851 + W1005 incidents cited as worked examples).
• W462 — test(landing): NEW landing-page tool-count drift-guard test (11 integers asserted across header/body/footer/docs/about; 1 pass).
• W844 — feat(build): auto-rotate _EXPECTED_CARD_SHA256 in dev/build_readme_counts.py (10 tests pass; closes W1308 manual-sync gap as a drive-by).
• W1038 — feat(yaml-loader): NEW extract_typed helper + validator kwarg follow-up (4 callsites migrated; 11 + 447 tests pass; cmd_alerts:961 == 0 clause clarification — NOT dead code).

Fixed — W1086-arc + Wave-B-TERMINAL + W478 + Pattern-1A-family batch

• W1084 — fix(ai-readiness): cmd_ai_readiness denominator-clamp probe-breaking fix — clamped to minimum-1 floor with insufficient-data disclosure (10 tests pass; W-number-collision target separate from the W1084 cmd_fitness arc).
• W478 — fix(tests): 4 SQLite fd-leak fixes in _make_db() test helpers — wrapped each in try/finally + explicit conn.close() (135 tests pass).
• W459 — refactor(mcp): normalize 17 MCP wrappers to description= kwarg for AST-walk consistency (2895 tests pass; carry-forward W449/W458/W459/W460 batch).
• W847 — fix(preflight): cmd_preflight UPPER-case scope clarification — W759 scope reduced 86% (30 → 4 sites).
• W759 — fix(envelope): 4 envelope-slot UPPER-case sites cleaned across cmd_preflight (13 tests pass; W762 cmd_preflight allowlist now empty).
• W1088 — fix(preflight): cmd_preflight _SEVERITY_ORDER lookup-miss belt-and-suspenders fix — lower-case canonical + UPPER-case aliases + .lower() at lookup (64 tests pass; W-number-collision target separate from the W1088 CI hardening arc).
• W851 BAIL — verify(tests): test_w596_confidence_level_rank_round_trip pre-existing failure, not reproducible in isolation; likely cross-worker warnings.resetwarnings() leak under xdist; captured for re-triage.

Changed — W1086-arc + Wave-B-TERMINAL batch

• SARIF dashboard family TERMINAL — the W1062 + W1062-followup + W1062-followup-2/3/4 fan-out + W1087 lint substitute arc closes cleanly. 12 wired emitters + 27 exempt (compound aggregators + thin advisories + invocation-scoped signals exempted by rationale) = 39 catalogued emitters pinned end-to-end. The W1062-followup-4 substitute-rather-than-wire recommendation realized.
• MCP outputSchema roadmap — Wave B TERMINAL closes the W767 5-wave specialized-schema propagation (CONSOLIDATE-19); Wave C1 opens the compat-profile arc at the env-var tier. Wave C2+ (roam mcp doctor probe surface) queues as the next MCP-roadmap milestone.
• Pattern-2 + Pattern-1A empty-state arc — closes at 8 detectors + 2 hard-cap commands sealed. Pattern-2 detector empty-state coverage now 8/8 effective sites; the 2 Pattern-1A hard-cap disclosure fixes (W1085 fingerprint + W1086 cut) follow the mirror-fix template established by the W1085 cmd_fitness SARIF advisory-plumb arc at CONSOLIDATE-17.
• Research-memo cluster — MCP outputSchema + compat-profile + FP-rate research memos now span 3 dated memos (one per consolidation pass since -18): MCP-OUTPUTSCHEMA-EVOLUTION (-18), DETECTOR-FP-RATE-METHODOLOGY (-19), MCP-COMPAT-PROFILE-ROADMAP (-20).

Added — Wave-B-TERMINAL + W794 + W1028 + W805 batch (post-CONSOLIDATE-18, 2026-05-16 /loop iteration N+19)

> ~18 completions since CONSOLIDATE-18 (Section 64). Wave B > TERMINAL milestone — the W767 5-wave outputSchema roadmap that > kicked off at CONSOLIDATE-18 closes cleanly across 13 MCP tools + > ~113 envelope-validation tests. Three themes carry the batch: > (a) Wave B TERMINAL — 13 MCP tools specialized across 5 > sub-ships (Wave B2 / B3 / B4 / B5-partial / B5b TERMINAL). > (b) MCP server card SEP-2127 readiness (W794) — icons[] > field across 4 .well-known path variants, carries the W792 + W793 > work to a clean SEP-2127-ready posture. (c) Pattern-2 > empty-state audit arc closure (W805) — 3 detectors migrated > (cmd_test_hermeticity + cmd_llm_smells + cmd_boundary) + 5 > followups captured + 1 real bug fixed (cmd_boundary SQL outside > with open_db block). Plus 3 stand-alone polish items > (W1061-followup-2 + W1008 carry-forward + the > DETECTOR-FP-RATE-METHODOLOGY research memo). Parallel in stature > to the W1255 architectural-decision-and-implementation arc that > CONSOLIDATE-16 carried.

• Wave-B2 — feat(mcp): specialized _SCHEMA_HEALTH + _SCHEMA_UNDERSTAND outputSchemas on roam_health + roam_understand wrappers (25 tests pass).
• Wave-B3 — feat(mcp): bundled _SCHEMA_ORACLE across 6 oracle wrappers (37 tests pass; single shared schema for the uniform oracle envelope shape).
• Wave-B4 — feat(mcp): specialized _SCHEMA_TIMELINE + _SCHEMA_TEST_IMPACT outputSchemas on roam_timeline + roam_test_impact wrappers (7 tests pass).
• Wave-B5-partial — feat(mcp): specialized _SCHEMA_AUDIT_TRAIL_VERIFY + _SCHEMA_DIAGNOSE outputSchemas (5 tests pass).
• Wave-B5b — feat(mcp): TERMINAL. Specialized _SCHEMA_FETCH_HANDLE + _SCHEMA_VALIDATE_PLAN + _SCHEMA_AUDIT_TRAIL_CONFORMANCE outputSchemas (39 tests pass). Closes the W767 5-wave outputSchema roadmap end-to-end across 13 MCP tools.
• W794 — feat(mcp-card): wire icons[] field across 4 .well-known path variants — mcp-server-card.json + .well-known/mcp-server-card + SEP-1649 mirror + SEP-2127 mirror (22 tests pass; SEP-2127 ready).
• W1028 — feat(formatter): _ALWAYS_PRESERVED_LIST_FIELDS expansion audit — 4 candidates marked DEFER + drift-guard test pinning the current set (162 tests pass; carry-forward CONSOLIDATE-17 → -18 → -19).
• W1008 — feat(formatter): surface list_counts top-level in strip_list_payloads so callers see how many items were stripped when --detail is off (234 tests pass; carry-forward W1000 drive-by).
• W1061-followup-2 — refactor(sarif): extract runtime_filter_disclosure() shared helper from 4 SARIF callers (cmd_smells + cmd_check_rules + cmd_taint + cmd_vulns); -36 LOC consolidation (17 tests pass).
• DETECTOR-FP-RATE-METHODOLOGY — docs(research): memo at dev/DETECTOR-FP-RATE-METHODOLOGY-2026-05-16.md (674 words, 12 sources cited; methodology for measuring detector false-positive rates). <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->

Fixed — Wave-B-TERMINAL + W794 + W1028 + W805 batch

• W805 — fix(detectors): Pattern-2 empty-state audit on cmd_test_hermeticity + cmd_llm_smells + cmd_boundary — all 3 detectors migrated to emit explicit partial_success: true on empty-state branches (13 tests pass; 5 followups captured as W805-followup-A/B/C/D/E).
• W805 real-bug fix — fix(boundary): cmd_boundary had a SQL execution block outside the with open_db context manager (resource-leak-on-error path); fixed inline during the W805 Pattern-2 migration.

Changed — Wave-B-TERMINAL batch

• MCP outputSchema roadmap — Wave B TERMINAL. The W767 5-wave specialized-schema propagation that kicked off with Wave B1 (CONSOLIDATE-18) lands on 13 MCP tools end-to-end (~113 cumulative envelope-validation tests pass). Wave C (compatibility profile + roam mcp doctor probe) queues as the next major MCP-roadmap milestone.
• MCP server card SEP-2127 readiness — W792 (3 .well-known path variants) + W793 (display_name → title rename) + W794 (icons[]) collectively land a clean SEP-2127-ready posture; W795 (_meta privacy posture stanza) remains BLOCKED on SEP-2127 upstream merge.
• Pattern-2 empty-state audit arc — closes at W805. The W802 → W836 sweep (CONSOLIDATE-17 + -18) covered the bulk of the detector roster; W805 adds the 3 remaining branches (cmd_test_hermeticity + cmd_llm_smells + cmd_boundary). Pattern-2 empty-state coverage on the detector roster is structurally complete; the 5 captured followups are surface-level disclosure-consistency polish.

Added — W1275-W1312-arc + Wave-B1 + sarif-disclosure batch (post-CONSOLIDATE-17, 2026-05-16 /loop iteration N+18)

> ~15 completions since CONSOLIDATE-17 (Section 63). Fast-follow- > through batch — 5 themes carry the dispatch tail: (a) Pattern-2c > carry-forward closures (W1275 / W1276-fix / W1277 / W1278a / W1309 > — the Pattern-2c CONSOLIDATE-16 → -17 → -18 chain closes cleanly, > Pattern-2c roster now 31/31 effective). (b) SARIF dashboard- > filtering trio (W1060 + W1061 + W1062 + 2 followups — OASIS 2.1.0 > § 3.51 + § 3.52 + properties.tags[] plumbed across cmd_smells + > cmd_check_rules + cmd_taint + cmd_vulns + cmd_health + cmd_complexity > + secrets emitter). (c) MCP outputSchema roadmap kickoff (W767 > inventory + Wave B1 specialized schemas on roam_impact + > roam_preflight + W1311 decorator normalization + W1312 redundancy > drops + the EVOLUTION research memo). (d) Pattern-1D file-substring > disclosure (W1309). (e) Pattern-3a severity widening (W1005 + > W1005-followup-A + W1007 agent_contract:[] preservation).

• W767 — docs(mcp): outputSchema inventory at dev/MCP-OUTPUTSCHEMA-INVENTORY-2026-05-16.md (57 core tools catalogued, 5-wave Wave B roadmap). <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• MCP-OUTPUTSCHEMA-EVOLUTION — docs(mcp): research memo at dev/MCP-OUTPUTSCHEMA-EVOLUTION-2026-05-16.md (Claude Code #25081 status shifted; 3-wave roadmap). <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• Wave-B1 — feat(mcp): specialized _SCHEMA_IMPACT + _SCHEMA_PREFLIGHT outputSchemas on roam_impact + roam_preflight wrappers (18 tests pass).
• W1060 — test(sarif): runtime-notifications activation tests on cmd_health + cmd_complexity (12 tests pass; cmd_doctor BAIL — no --sarif flag).
• W1061 — feat(sarif): ruleConfigurationOverrides[] on cmd_smells (OASIS 2.1.0 § 3.51 compliant, default-off, 38 tests pass).
• W1061-followup — feat(sarif): extend ruleConfigurationOverrides[] + add notificationConfigurationOverrides[] on cmd_check_rules + cmd_taint + cmd_vulns (11 tests pass).
• W1062 — feat(sarif): result.properties.tags[] on taint + vulns + audit-trail-conformance (21 tests pass).
• W1062-followup — feat(sarif): result.properties.tags[] on secrets_to_sarif emitter (60 tests pass).
• W1278a — feat(test-scaffold): Convention-c migration for cmd_test_scaffold (Pattern-2c roster now 31/31 effective; 22 tests pass).
• W1309 — feat(test-scaffold): Pattern-1D file-substring resolution: "file_substring" enum disclosure on cmd_test_scaffold (31 tests pass).

Fixed — W1275-W1312-arc + Wave-B1 + sarif-disclosure batch

• W1005 — fix(smells): widen cmd_smells --min-severity from 3-tier to W547 7-tier canonical via severity_rank lookup (Pattern 3a; 236 tests pass).
• W1005-followup-A — fix(llm-smells): parallel 7-tier widening on cmd_llm_smells --min-severity (Pattern 3a vocabulary uniformity; 2 tests pass).
• W1007 — fix(formatter): strip_list_payloads now preserves agent_contract:[] without --detail (closes W1006 audit finding; 89 tests pass).
• W1275 — fix(tests): harden 3 remaining dogfood-brittle assertions in test_validate_plan.py (carry-forward CONSOLIDATE-16 → -17 → closed at -18; 27 tests pass).
• W1276-fix — NO-OP verified: fix already landed during W1272 Pattern-2c batch (structural verification only).
• W1277 — fix(impact): restore auto_log provenance on cmd_impact unresolved-attempt path so roam replay narrates the unresolved-attempt event (W1272 standardization had removed it; 8 tests pass).
• W1311 — refactor(mcp): normalize 5 oracle multi-line @_tool( decorators so inventory script AST-walks cleanly (-37 LOC; 131 tests pass).
• W1312 — refactor(mcp): drop 3 redundant output_schema=_ENVELOPE_SCHEMA declarations + queue 2 for Wave B (142 tests pass).
• sarif-disclosure — fix(docs): cmd_boundary + cmd_compatibility + cmd_test_hermeticity docstrings add --sarif flag disclosure (closes CI-blocking drift-guard gap; 103 tests pass).

Changed — W1275-W1312-arc batch

• Pattern-2c roster — effective 31/31 closed (30 from CONSOLIDATE-15 + W1278a at CONSOLIDATE-18). The CONSOLIDATE-16 → -17 → -18 carry-forward chain (W1275 / W1276-fix / W1277 / W1278) closes cleanly.
• SARIF emitter coverage — dashboard-filtering plumb (rule + notification config overrides + result tags) now lands across 7 emitter surfaces. The OASIS 2.1.0 § 3.51 + § 3.52 plumb is structurally complete on the high-signal-tier emitters.
• MCP outputSchema roadmap — Wave A (default _ENVELOPE_SCHEMA) inventory catalogued; Wave B specialized-schema propagation kicked off with Wave B1 on the 2 highest-frequency wrappers (roam_impact + roam_preflight).

Added — W1284-W1308 batch (post-CONSOLIDATE-16, 2026-05-16 /loop iteration N+17)

> ~25 completions since CONSOLIDATE-16 (Section 62). Post-v13.2-release > hardening batch — first ~25 W#s after the release-merge land cleanly > without re-opening flagship arcs. Four themes carry the batch: (a) init / > cold-start UX fixes (W1288 / W1289 / W1290 / W1291). (b) SARIF advisory- > warning plumb carry-forward bundle (W1084 + W1113 + W1114 + W1115 in one > commit + W1236 chore drop of orphan emitters). (c) CGA edge-bundle > stability + post-merge CI hardening (W1285 / W1284-G3 / W1286 / W1287 + > W1297-W1302 / W1303-W1305). (d) MCP card v13.2 sync + CI infrastructure > (W1306 / W1307 / W1308 / W1088 / W1089).

• W1287 — feat(test-hermeticity): non-hermetic test detector (new smell-detector family member).
• W1292 — docs(plugin): close Gap 3 — 3-hook copy-fork template under dev/example-plugin/.

Fixed — W1284-W1308 batch

• W1084 — feat(sarif): plumb advisory warnings into cmd_fitness emitter (shipped via 96d31bd0 with W1113/W1114/W1115).
• W1113 — feat(sarif): plumb advisory warnings into cmd_flag_dead emitter.
• W1114 — feat(sarif): plumb advisory warnings into cmd_rules emitter.
• W1115 — feat(sarif): plumb advisory warnings into cmd_health emitter.
• W1236 — chore(sarif): drop orphan breaking + conventions emitters (no registry consumers post-W1232).
• W1284-G3 — fix(relations): SFC synthetic-component anchor for module-scope imports.
• W1285 — fix(cga): edge_bundle_digest sort-stability — append id tiebreaker.
• W1286 — perf(clones): language allowlist on candidate fetch (~3-5× candidate-set reduction on multi-language repos).
• W1288 — fix(init): drop misleading "Health: N/100" banner on cold-start (was always 100/100 on 0-symbol corpus).
• W1289 — fix(mcp-status): canonical Pattern-1A envelope on fastmcp import fail (was raw ImportError).
• W1290 — fix(surface): AST-derived mcp_tool_count survives [mcp] extras gap (cold install without fastmcp).
• W1291 — test(init): regression — cmd_init must not self-recommend roam init (silent advisory-loop prevention).
• W1297 (follow-up) — fix(format): ruff format 6 drift-guard test files left unformatted by v13.2 merge.
• W1298-W1302 — fix: 6 CI failures on main — drift-guard re-pins + CGA dirty_tree=true propagation through pr-bundle integration tests.
• W1303-W1305 — fix: doc-hygiene drift + ruff I001 + W792 server-card mirror sync.
• W1306 — fix: server.json + changelog.html v13.2 catchup (landing page was pinned at v13.1).
• W1307 — fix(test): bump _EXPECTED_CARD_SHA256 to v13.2 card digest.
• W1308 — fix: LF-normalize 3 MCP card files (mcp-server-card.json + .well-known/ mirror + SEP-1649 variant); content-sync resolves CRLF-drift footgun on Windows checkouts.

Changed — CI infrastructure

• W1088 — ci: SHA-pin credential-bearing third-party actions (immutable SHAs on publish.yml + release pipeline).
• W1089 — ci(publish): replace sleep-45 smoke-job with retry-backoff (intermittent v13.0 / v13.1 publish-failure root cause).

[13.2] — 2026-05-16

Highlights — wip/v13.2-session-2026-05-16 ship (W1255-W1297)

The 2026-05-16 session that landed:

• W1272 Pattern-2c standardization — 8 commands now emit canonical exit-0 + "not found" envelope on unresolved paths (impact, preflight, trace, testmap, context, safe-delete, split, why).
• W1255 config-hash producer wire-up — .roam-rules.yml + .roam/constitution.yml + .roam/control-map.yml canonical paths stamped onto every run; evidence/config_hashes.py substrate + ledger producer wire-up. W1253 unblocked. vsa.py already CONSUMES the hashes so VSA attestation benefits immediately.
• W210 W211 producer coverage — pr-replay path achieves 7-complete + 1-partial of 8 evidence questions on roam-code itself; remaining gap (Q8 approvals) explicitly marked via producer_not_available.
• 30+ CI hardening fixes (W1281-W1297) — TestBatchSearch cold-start guard, mode-classification taxonomy, eight-questions GitHub Actions skip, README v11→v13.2 narrative refresh, .githooks executable bit, ~17 dogfood-corpus-dependent tests now centrally skipped via conftest.py collection hook, LAW-4 verdict refresh on minimap, Pattern-2c test assertion updates, mode-policy W1288 (why-fail / why-slow / workflow classified read_only), and the W762 cmd_preflight line-allowlist refresh.
• Documentation — 224 MCP tools (was: 137) surfaced; v13.2 README agentic-assurance frame replaces v11 roadmap section.

Added — W1255-W1278 batch (post-CONSOLIDATE-15, 2026-05-16 /loop iteration N+16)

> 7+ completions since CONSOLIDATE-15 (Section 61). The follow- > through batch after the Pattern-2c 30/30 terminal landed. The > MAJOR load-bearing milestone: the W1255 architectural > decision (Option (a) "Keep top-level + add siblings") landed AND > shipped within the same consolidation window — .roam-rules.yml > (root) + .roam/constitution.yml (existing) + .roam/control-map.yml > (new) are the canonical config paths, and src/roam/evidence/config_hashes.py > (84 LOC, NEW) + ledger.py stamping at start_run (+18 LOC) wire > the producer side end-to-end. Side benefit: vsa.py already > CONSUMES constitution_hash + rules_config_hash at lines > 281-296 — producer wire-up immediately benefits VSA attestation > with zero further code change. W1253 unblocked. Plus the > W1272 Pattern-2c unresolved-path standardization milestone: > 8 commands (cmd_impact + 6 helper-callers + cmd_preflight > already-compliant pin) now emit the canonical Convention-c > unresolved-path shape — exit code 0 on unresolved across all 8. > Five themes: (1) W1255 architectural decision recorded + IMPL > shipped — Cranot picked Option (a); config_hashes.py substrate > + ledger.py producer wire-up + CLAUDE.md doc landed inside the > same window. 11 new tests + 101 in-scope tests pass; hash-stability > preserved. (2) W1272 Pattern-2c unresolved-path standardization > SHIPPED — 8-command Convention-c bulk migration (78+105+27+51 > tests pass; zero regressions; exit-code-0-on-unresolved across all > 8). The post-Pattern-2c-terminal follow-up arc surfaced at > CONSOLIDATE-15 (W1268-audit) lands within the next consolidation > window. (3) W1273 test_validate_plan dogfood-brittleness fix > SHIPPED — 3 tests hardened (cold-start-guard bypass + > _vp_blast_radius stubbing); 27/27 tests pass. (4) Drive-by > captures — W1275 (3 remaining dogfood-brittle tests in > test_validate_plan.py) + W1276 (test_impact_auto_logs_not_found_path > RECLASSIFIED → W1272-expected-failing; in flight as W1276-fix) + > W1277 (replay-narration provenance for unresolved-path attempts > — auto_log removed from cmd_impact; signal-loss risk) + W1278 > (audit 3 remaining symbol_not_found callers — cmd_test_scaffold > / cmd_plan_refactor / cmd_guard). (5) Lockstep consolidation > discipline — the every-~8-completions /loop rule fires; the > follow-through batch lands cleanly without the multi-arc-terminal > volume of CONSOLIDATE-14 / CONSOLIDATE-15.

• W1255 architectural decision recorded AND shipped within the same window. Cranot picked Option (a) "Keep top-level + add siblings". Canonical paths: .roam-rules.yml (root) + .roam/constitution.yml (existing) + .roam/control-map.yml (new). W1255-IMPL shipped: src/roam/evidence/config_hashes.py (84 LOC, NEW) + ledger.py stamping at start_run (+18 LOC) + CLAUDE.md doc (+17 LOC). 11 new tests + 101 in-scope tests pass. Hash-stability preserved. Side benefit: src/roam/evidence/vsa.py already CONSUMES constitution_hash + rules_config_hash at lines 281-296 — the producer wire-up immediately benefits VSA attestation with zero further code change. W1253 unblocked (the W1255 decision was the only blocker remaining on W1253). The architectural-decision-and-implementation arc fitting inside a single consolidation window is the canonical fast-path for substrate-first sequencing — decision captured at CONSOLIDATE-14, decision-locked + implemented within the post-CONSOLIDATE-15 follow-through batch.
• W1272 Pattern-2c unresolved-path standardization SHIPPED — 8-command Convention-c bulk migration. Post-CONSOLIDATE-15 follow-through arc that closes the W1268-audit-captured 5-way unresolved-path divergence at the consumer level. cmd_impact + 6 helper-callers (cmd_dead / cmd_safe_delete / cmd_closure / cmd_symbol / cmd_hover / cmd_pytest_fixtures — illustrative; see commit for the full list) + cmd_preflight already-compliant pin now emit the canonical Convention-c unresolved-path shape. 78+105+27+51 tests pass. Zero regressions. Exit code 0 on unresolved across all 8 commands (was previously divergent — some emitted exit-2, some exit-5, some exit-0 with partial_success: true). The Pattern-2c terminal at 30/30 (CONSOLIDATE-15) addressed the disclosure of degraded resolution; W1272 addresses the shape consistency of the unresolved-path branch across consumers. Together: Pattern-2c is now both disclosure-complete AND shape-uniform.
• W1273 — test_validate_plan dogfood-brittleness fix SHIPPED. The W1271-audit / W1273-capture / W1274-fix arc from CONSOLIDATE-15 covered the test_visualize stale-assertion case; W1273 proper covered the remaining test_validate_plan cases. 3 tests hardened: cold-start-guard bypass + _vp_blast_radius stubbing applied. 27/27 tests pass. The capture-to-fix arc spans CONSOLIDATE-15 → CONSOLIDATE-16 — clean cross-session follow-through on the dogfood-brittleness surface.
• Drive-by captures during this consolidation window (4 captures). W1275 (harden 3 remaining dogfood-brittle tests in test_validate_plan.py — partial W1273 follow-up). W1276 (test_impact_auto_logs_not_found_path RECLASSIFIED → W1272-expected-failing; test needs update; in flight as W1276-fix). W1277 (restore replay-narration provenance for unresolved-path attempts; auto_log was removed from cmd_impact during W1272 standardization — there's a signal-loss risk on the replay-narration surface that wants explicit recovery). W1278 (audit 3 remaining symbol_not_found callers — cmd_test_scaffold / cmd_plan_refactor / cmd_guard — for Convention-c alignment; the W1272 bulk migration touched 8 of 11 known callers, the remaining 3 want an audit before bulk-migration).

In flight — W1255-W1278 batch (parallel dispatches not yet on disk)

• W1253 — was unblocked by W1255-IMPL landing; the next-session dispatch will pick up.
• W1276-fix — test_impact_auto_logs_not_found_path test-needs-update (W1272 follow-up; expected-failing under the new W1272 exit-code-0 contract).

Added — W1245-W1274 batch (post-CONSOLIDATE-14, 2026-05-16 /loop iteration N+15)

> ~20 completions since CONSOLIDATE-14 (Section 60) — the largest > cumulative consolidation since the W1175-RESEARCH propagation arc > mid-points. The MAJOR milestone: Pattern-2c propagation arc > COMPLETE at 30/30 sites. The W1233-audit roster (originally 38 > sites) resolved to 30 real true-positives once the W1267-audit > filtered out two W1233-audit false positives (cmd_hotspots / > cmd_smells lacked real find_symbol callsites). Wave 1 quartet > (W1242 + W1243 + W1244 + W1248 — CONSOLIDATE-14) + cmd_annotate > (W324 origin template) + W1245 batches 1-4 covering 20 SHIP + > 2 BAIL (22 cmd_*.py visited, 20 disclosure-covered + 2 false- > positive BAILs) closed every remaining real Pattern-2c site. > Both terminal > arcs are now CLOSED: the SARIF SHIP/SKIP-disclosure 196 → 0 > propagation arc reached terminal at CONSOLIDATE-14; the Pattern-2c > 30/30 arc reaches terminal at CONSOLIDATE-15. The agentic-assurance > substrate now spans producer (W1234 evidence_stale + earlier W210 > packet substrate) + consumer (W1262 doctor/diff stale banner) + > attestation (W37x CGA + W377 permit collector) — all three axes > structurally complete. Six themes: (1) Pattern-2c bulk > completion — W1245 batches 1-4 (22 SHIP across cmd_dead + > cmd_safe_delete + cmd_closure + cmd_symbol + cmd_hover + > cmd_pytest_fixtures + cmd_plan + cmd_context + cmd_relate + > cmd_why + cmd_visualize + cmd_invariants + cmd_testmap + > cmd_affected_tests + cmd_guard + cmd_metrics + > cmd_plan_refactor + cmd_pr_bundle + cmd_safe_zones + > cmd_test_scaffold) + 2 BAIL on W1233-audit false positives > (cmd_hotspots / cmd_smells — no real find_symbol callsite). > Plus cmd_annotate origin template (W324) accounted in the > 30-tally. (2) Pattern-2c family extensions — W1250 helper > docstring (collision-pattern documented); W1270 helper reserved- > key warning (Pattern-2 silent-drop fix at substrate level; first > real-world use in W1245-batch-4 cmd_safe_zones); W1268-audit > surfaced 5-way unresolved-path divergence captured as W1272; > W1271-audit surfaced test_validate_plan dogfood-brittleness > captured as W1273; W1273-fix → W1274 stale-assertion fix in > test_visualize; W1265 docstring at vsa.py:133 (W1264 follow-up). > (3) Evidence/W210 extensions — W1262 doctor/diff stale-evidence > banner (consumer-side wire-up of W1234 evidence_stale producer); > W1266 completeness_compat shared module hoist (-180 LOC duplicate > helpers + 205 LOC shared; W1262 drive-by). (4) Per-kind version > stamps — W1256 cmd_vibe_check per-pattern version stamps (10 > patterns); W1269 cmd_smells per-kind version stamps (7 patterns > wired). (5) Audit closures — W1267 audit corrected the 34-site > Pattern-2c list to 30 real true-positives by filtering the two > W1233-audit false positives. (6) CONSOLIDATE pause — natural > stopping point after Wave 2 batch-4 lands; no in-flight dispatches > at consolidation time. (Tally arithmetic: 20 SHIP across W1245 > batches 1-4 + 4 Wave-1 from CONSOLIDATE-14 + 1 W324 cmd_annotate > origin + 5 already covered upstream / earlier = 30 real > Pattern-2c sites disclosure-covered; the W1233-audit 38-site > original count was inflated by 2 false positives + ~6 duplicates > already covered by earlier substrates.)

• Pattern-2c propagation arc COMPLETE at 30/30 sites (W1245 batches 1-4). The MAJOR load-bearing milestone: every real Pattern-2c site in the W1233-audit enumeration now disclosures resolution-tier via the W1241 resolution_disclosure() helper substrate. Batch breakdown. W1245-batch-1 (3 SHIP — cmd_dead + cmd_safe_delete + cmd_closure; 2 BAIL — cmd_hotspots / cmd_smells were W1233-audit false positives, no real find_symbol callsites). W1245-batch-2 (5 SHIP — cmd_symbol + cmd_hover + cmd_pytest_fixtures + cmd_plan + cmd_context). W1245-batch-3 (5 SHIP — cmd_relate + cmd_why + cmd_visualize + cmd_invariants + cmd_testmap). W1245-batch-4 (7 SHIP — cmd_affected_tests + cmd_guard + cmd_metrics + cmd_plan_refactor + cmd_pr_bundle + cmd_safe_zones + cmd_test_scaffold). Plus the Wave-1 quartet from CONSOLIDATE-14 (W1242 cmd_impact + W1243 cmd_preflight + W1244 cmd_diagnose + W1248 cmd_trace) and the W324 cmd_annotate origin template — together 30/30 real Pattern-2c sites. W1267-audit corrected the W1233-audit roster (originally 38 sites; surfaced 2 false positives + ~6 duplicates already covered upstream → 30 real true-positives). The W1249 substrate refactor's ~3× LOC simplification per consumer made batches-2/3/4 tractable at ~7 sites per dispatch versus Wave-1's ~1-site-per-dispatch cadence. Hash-stability invariant held throughout: every adoption byte-stable for symbol-tier (exact-match) envelopes; only the partial-success branches emit new field bytes. The propagation arc is structurally complete — no surviving Pattern-2c gaps remain in the cmd_*.py surface.
• Pattern-2c family extensions (W1250 + W1270 + W1268-audit + W1271-audit + W1273-fix + W1274 + W1265). W1250 expanded the resolution_disclosure() helper docstring with the W324 cmd_annotate template precedent + W1241 substrate-first sequencing + collision-pattern documentation (what happens when two resolvers race on the same logical symbol). W1270 added the helper's reserved-key warning surface — the substrate now flags Pattern-2 silent-drop when a downstream caller tries to override a reserved envelope key; the first real-world use landed in W1245-batch-4 cmd_safe_zones, where a name-collision between the resolution-tier output and a domain-specific partial_success field was caught at substrate layer rather than silently dropped. W1268-audit surfaced a 5-way unresolved-path divergence across the Pattern-2c consumer family (each cmd hand-rolled its own degraded-resolution path) — captured as W1272 for bundled standardization (10 cmd_*.py / ~150 LOC). W1271-audit surfaced test_validate_plan dogfood-brittleness (assertion couples to an unstable transient hash) — captured as W1273. W1273-fix shipped as W1274 (test_visualize stale-assertion fix; ~10-20 LOC). W1265 added a load-bearing docstring at src/roam/evidence/vsa.py:133 (W1264 follow-up — surfaced during W1262 stale-banner wiring). The Pattern-2c family's substrate-first sequencing precedent (substrate → Wave-1 → substrate refactor → bulk migration) is now the canonical playbook for closed-vocab propagation arcs.
• Evidence/W210 extensions (W1262 + W1266). W1262 landed the consumer-side wire-up of the W1234 evidence_stale producer — roam doctor and roam diff now surface a "stale evidence" banner when the W1234-emitted evidence_stale: true field appears on the consumed packet (closes the W1254 in-flight dispatch from CONSOLIDATE-14). W1266 hoisted evidence_completeness_compat helpers into a shared module (-180 LOC of duplicate helpers across cmd_doctor / cmd_diff / cmd_critique / 3 sibling sites + 205 LOC shared module) — drive-by from W1262 stale-banner wiring; W1266 is the substrate-first-sequencing exemplar applied to evidence completeness checks. The agentic-assurance substrate now spans all three axes — producer + consumer + attestation — structurally complete. The W1234 producer (CONSOLIDATE-14) emits; the W1262 consumer (CONSOLIDATE-15) surfaces; the W37x CGA + W377 permit-collector attest. The W210 packet-layer Pattern-2 variant-2f family is now end-to-end live; no producer/consumer/attestation gap survives on the evidence-staleness axis.
• Per-kind version stamps — W1256 (vibe-check) + W1269 (smells). W1256 added per-pattern detector version stamps to cmd_vibe_check's 10 AI-rot patterns (each pattern's emission now carries its own pattern_version field — agents can detect a pattern's signal shape changed without forcing the whole detector version to bump). W1269 wired per-kind version stamps for the 7 cmd_smells patterns that still shared the composite SMELLS_DETECTOR_VERSION fallback (W870 vintage — 7/24 detectors had per-id stamps post-W870; W1269 closes the 7 most-touched of the remaining 17). Both ships are byte-stable additive — no persisted finding rows touched; the new version fields land on the FindingRecord envelope alongside (not replacing) the composite detector version. Together W1256 + W1269 close the version-stamp gap for the two largest pattern-family detectors in the catalog (vibe-check + smells).
• Audit closures — W1267. W1267 corrected the W1233-audit roster from 34 sites to 30 real true-positives. The audit found that cmd_hotspots and cmd_smells listed in W1233-audit Wave-2 batch-1 lacked a real find_symbol() callsite — both commands resolve their inputs via the rule-engine path rather than the symbol-resolver path, so the Pattern-2c disclosure shape doesn't apply (no degraded resolution to disclose). The two BAIL outcomes (W1245-batch-1 BAILs) are recorded with structured rationale; W1233-audit's 38-site original count resolved cleanly to 30 real sites once W1267 + duplicates-already-covered cross-referenced.

In flight — W1245-W1274 batch (parallel dispatches not yet on disk)

• (none — natural pause point after W1245-batch-4 lands; no parallel dispatches active at consolidation time)

Added — W1242-W1259 batch (post-CONSOLIDATE-13, 2026-05-16 /loop iteration N+14)

> ~15 completions since CONSOLIDATE-13 (Section 59). Six themes: > (1) Pattern-2c family enablement — Wave 1 quartet landed — > W1242 cmd_impact + W1243 cmd_preflight + W1244 cmd_diagnose + > W1248 cmd_trace adopted the W1241 resolution_disclosure() helper > substrate at the find_symbol() / find_symbol_id() call sites. > The four flagship commands now surface which tier of the resolver > succeeded (symbol / file / fuzzy / unresolved) + a > partial_success flag set on any non-symbol resolution. Wave 1 > finishes the highest-traffic Pattern-2c sites first, matching the > W1192/W1195 SARIF SHIP sequencing pattern. (2) Pattern-2c > substrate refactor (W1249) — hoisted find_symbol tier-stamping > into the substrate helper, eliminating ~100 LOC of duplicate > _detect_resolution_tier helpers across the four Wave-1 flagships. > The W1249 refactor unblocks W1245-batch-1 (Wave 2 first 5 sites) at > ~3× LOC simplification per consumer — without it, each Wave-2 > adoption would carry ~25 LOC of boilerplate the substrate now > absorbs. (3) Wave 16 SKIP-disclosure landed — _KNOWN_MISSING > 17 → 0 — 17 docstrings shipped across the remaining Bucket B > long-tail (cmd_debt + cmd_entry_points + cmd_guard + cmd_map > + cmd_metrics + cmd_path_coverage + cmd_patterns + > cmd_plan_refactor + cmd_pytest_fixtures + cmd_risk + > cmd_safe_delete + cmd_safe_zones + cmd_simulate_departure + > cmd_suggest_refactoring + cmd_testmap + cmd_why_slow + > cmd_ws). The 196 → 0 propagation arc is now fully closed — > 196 commands audited, 196 commands disclosure-covered (179 > SKIP-disclosure docstrings + 17 SARIF SHIP emitters across > CONSOLIDATEs 4 → 14). The W1175-RESEARCH long-tail roster > exhausted; arc terminal. (4) Evidence/W210 wire-up — W1234 > shipped the evidence_stale producer (W210 packet-layer Pattern-2 > variant-2f); W1254 (consumer) in flight at consolidation time; W1253 > BAIL surfaced W1255 architectural prerequisite (no upstream packet > exists to mark stale → captured as architectural decision pending). > (5) State-vocab substrate (W1235) — _STATE_FAMILY_ALIASES > registry landed at substrate level for state-name normalization > across closed-vocab Pattern-2g sites. (6) SARIF rule rename > (W1232) — flag-constant-default rule renamed to flag-suspect > per W1226 SHIP scope-discipline follow-up, aligning the > cmd_flag_dead namespace closer to W1227/W1229 naming convention. > Plus three CLAUDE.md doc-drift refreshes (W1247 module-local SARIF > convention + W1252 findings-registry decision + W1258+W1259 > 16 → 26 detector count + emit_finding(conn, record) API name) > and two new research memos (dev/PATTERN-2-EVOLUTION-2026-05-16.md > <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> > already shipped in CONSOLIDATE-13; dev/DETECTOR-FP-RATE-BENCHMARKS-2026-05-16.md > <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> > 773 LOC false-positive rate benchmarks across the 26 emitting > detectors).

• Pattern-2c Wave 1 quartet landed (W1242 + W1243 + W1244 + W1248). Four flagship commands — cmd_impact (W1242) / cmd_preflight (W1243) / cmd_diagnose (W1244) / cmd_trace (W1248) — consume the W1241 resolution_disclosure() helper substrate at the find_symbol() / find_symbol_id() call sites. Envelopes now carry resolution (closed-enum: symbol / file / fuzzy / unresolved) + partial_success: true on any non-symbol resolution. Wave 1 covers the four highest-traffic Pattern-2c sites first per the W1192/W1195 SARIF SHIP sequencing pattern; Wave 2 (W1245-batch-1 in flight; 5 sites) and Wave 2/3 long-tail (29 remaining sites) carry forward to next session. The four Wave-1 commands are the exact ones that agents call most frequently in pre-edit workflows (roam impact for blast radius; roam preflight for gate-before-edit; roam diagnose for root-cause ranking; roam trace for symbolic path traversal) — so the user-facing partial-success disclosure lands in the highest-leverage call sites first.
• Pattern-2c substrate refactor (W1249). Hoisted the find_symbol tier-stamping into the canonical resolution_disclosure() helper at src/roam/output/formatter.py:1263, eliminating ~100 LOC of duplicate _detect_resolution_tier helpers across the four Wave-1 flagships (cmd_impact / cmd_preflight / cmd_diagnose / cmd_trace). The W1249 refactor unblocks W1245-batch-1 (Wave 2 first 5 sites — cmd_hotspots / cmd_smells / cmd_dead / cmd_safe_delete / cmd_closure) at ~3× LOC simplification per consumer — without it, each Wave-2 adoption would carry ~25 LOC of boilerplate that the substrate now absorbs. The refactor preserves byte-stable envelopes for the Wave-1 quartet (verified per-command via hash-comparison on representative JSON outputs) — no detector output bytes moved.
• Wave 16 SKIP-disclosure — _KNOWN_MISSING 17 → 0; the 196→0 propagation arc is fully closed. 17 SKIP-disclosure docstrings landed across the remaining Bucket B long-tail: cmd_debt + cmd_entry_points + cmd_guard + cmd_map + cmd_metrics + cmd_path_coverage + cmd_patterns + cmd_plan_refactor + cmd_pytest_fixtures + cmd_risk + cmd_safe_delete + cmd_safe_zones + cmd_simulate_departure + cmd_suggest_refactoring + cmd_testmap + cmd_why_slow + cmd_ws. The 196→0 propagation arc is terminal — 196 commands audited, 196 commands disclosure-covered (179 SKIP-disclosure docstrings + 17 SARIF SHIP emitters across CONSOLIDATEs 4 → 14). The W1175-RESEARCH long-tail roster has now been fully exhausted; no surviving _KNOWN_MISSING entries remain. The arc spanned W1146 → W1259 across 11 CONSOLIDATE waves and ~30 sessions; Wave 16 is the largest single SKIP-disclosure wave since Wave 14b (CONSOLIDATE-12, 22 docstrings) at 17 docstrings — closing the long-tail with 0 BAILs and 0 reclassifications, both inverse-drift guards (tests/test_known_missing_pin_is_current + tests/test_pattern_2_propagation_coverage) green at consolidation time.
• Evidence/W210 wire-up (W1234 + W1253 BAIL + W1254 in flight). W1234 landed the evidence_stale producer for the W210 packet-layer Pattern-2 variant-2f — the field is now populated upstream when the evidence-compiler detects time-skew between context_read_at / edits_started_at / edits_completed_at and the current commit. W1254 (consumer-side: the report renderer + projection layers consume evidence_stale to surface a "stale evidence" banner) dispatched in parallel at consolidation time. W1253 BAIL surfaced an architectural prerequisite — the pr-bundle emit path cannot mark a packet stale before any packet exists (no upstream packet to mark) — captured as W1255 architectural decision pending. The BAIL-and-capture discipline (Pattern-3b reclassification arc precedent) applies cleanly: a discovered prerequisite gap becomes a captured architectural decision rather than a silent no-op.
• State-vocab substrate (W1235). Landed the _STATE_FAMILY_ALIASES registry at substrate level for state-name normalization across closed-vocab Pattern-2g sites. Captures the W1077/W1080 structured_unknown_filter precedent — closed-vocabulary state-name divergence ("idle" vs "waiting" vs "pending" etc.) gets canonicalized at the substrate boundary rather than each detector hand-rolling its own alias table. The registry is the substrate-first equivalent of the W1018 YAML loader substrate (W965-CONSOLIDATE) and the W1241 resolution_disclosure() helper (CONSOLIDATE-13). 45-site bulk migration (W1251) captured for next session — same substrate-first sequencing.
• SARIF rule rename (W1232). Renamed the flag-constant-default rule under the flag- namespace to flag-suspect per the W1226 SHIP scope-discipline follow-up. The rename aligns cmd_flag_dead's closed-enum rule set closer to W1227 / W1229 naming convention — flag- rules now express "what the SARIF row claims about the flag" (flag-staleness / flag-single-reference / flag-suspect) rather than mixing claims with antecedent conditions (flag-constant-default named the cause rather than the consequence). Hash-stable for the rule-set rename via the SARIF wrapper's rule_id` field — no persisted finding rows touched.
• CLAUDE.md doc-drift refresh (W1247 + W1252 + W1258 + W1259). Four small but load-bearing CLAUDE.md updates: W1247 added the module-local SARIF convention note (_to_sarif() helpers live in the cmd module per SHIP emitter — not centralized — per W1236-audit BENIGN verdict); W1252 captured the findings-registry decision (emit_finding(conn, record) as canonical API name — supersedes the older findings_store.persist(...) snake_case spelling that drifted into early docs); W1258 + W1259 refreshed the detector count from "16 detectors persist findings" (W146 vintage) to "26 detectors persist findings as of 2026-05-16" with the 10 newly-emitting detectors enumerated (critique, doctor, fan, fingerprint, health, llm-smells, etc. — predominantly aggregator / consumer commands that re-emit derived findings from upstream detectors).
• Research memo — dev/DETECTOR-FP-RATE-BENCHMARKS-2026-05-16.md (773 LOC). False-positive rate benchmarks across the 26 emitting detectors. Reference for the next per-detector confidence-tier tuning pass (W1256 captured for next session). Companion to dev/PATTERN-2-EVOLUTION-2026-05-16.md (884 LOC, shipped CONSOLIDATE-13) — together the two memos cover the W2026-05-16 detector-quality landscape. No source changes; sequencing decisions are Cranot's. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• Drift-guard remediation (W1239 + W1240) — hygiene clean. W1239 (drift-guard hygiene follow-up A from W1231-audit) shipped — stale audit assertion cleaned up. W1240 (drift-guard hygiene follow-up B) shipped — BACKLOG.md table-of-pendings cleaned to match the post-CONSOLIDATE-13 ground truth.

Audits / verdicts — W1242-W1259 batch

• W1230 audit — cmd_test_gaps SKIP confirm (re-verification). Originally raised CONSOLIDATE-9 (W1202), re-opened CONSOLIDATE-13 by Wave 14b docstring landing, confirmed CONSOLIDATE-14: REPORT-not-detector pattern (no per-location FindingRecord persistence). SKIP-disclosure docstring stays in place. The verdict has now been re-affirmed across three consecutive consolidation passes — strongest signal yet that the classification is stable. Drive-by drift note: BACKLOG.md "Outstanding from older rosters" still lists cmd_test_gaps in stale carry-forward rosters in earlier CONSOLIDATE-12/13 sections — surfaced for cleanup but not load-bearing.
• W1231 audit — drift-guard triage (5-failure) — CLOSED. Two FIXED (W1237 cmd_risk edge-kind canonicalize + W1238 catalog/detectors.py bare-except migration — both landed CONSOLIDATE-13); two follow-ups SHIPPED CONSOLIDATE-14 (W1239 + W1240 hygiene clean).
• W1233 audit — 38 Pattern-2c sites enumeration — Wave 1 closed (quartet shipped CONSOLIDATE-14); Wave 2 batch-1 in flight; remaining 29 sites captured for Wave 2 batch-2+.
• W1236 audit — SARIF helper convention sweep — VERDICT BENIGN. Module-local SARIF convention consistent across all 37 emitters; no substrate canonicalization needed. W1247 doc-pass landed.
• W1246 audit — cmd_trace.find_symbol_id Pattern-2c — VERDICT NON-COMPLIANT (CONSOLIDATE-13). Captured as W1248; shipped CONSOLIDATE-14.
• W1257 audit in flight — ~45-site state-vocab adoption sweep (consumer-side of W1235 _STATE_FAMILY_ALIASES registry). Dispatched in parallel at consolidation time; non-CHANGELOG-touching.

In flight — W1242-W1259 batch (parallel dispatches not yet on disk)

• W1254 — evidence_stale consumer wire-up. Dispatched at consolidation time. The W210 packet-layer consumer-side: report renderer + projection layers consume the evidence_stale field W1234 populates and surface a "stale evidence" banner.
• W1245-batch-1 — Pattern-2c Wave 2 adoption (5 sites). Dispatched at consolidation time. cmd_hotspots / cmd_smells / cmd_dead / cmd_safe_delete / cmd_closure — the first 5 of the 10 Wave-2 sites surfaced by W1233-audit. W1249's substrate refactor cut per-consumer LOC ~3×, making this batch tractable in a single dispatch.
• W1257-audit — state-vocab adoption sweep (45 sites). Dispatched at consolidation time. Consumer-side of the W1235 _STATE_FAMILY_ALIASES registry; same pattern as the W1233-audit dispatch for Pattern-2c.

Added — W1226-W1248 batch (post-CONSOLIDATE-12, 2026-05-16 /loop iteration N+13)

> ~13 completions since CONSOLIDATE-12 (Section 58). Four themes: > (1) SARIF SHIP family grew from 34 to 37 emitters in a single > post-CONSOLIDATE-12 window — W1226 cmd_flag_dead (35th, three closed- > enum rules under the flag-* namespace: flag-staleness / > flag-single-reference / flag-constant-default; staleness-banded > per-result level with a warning ceiling — heuristic detector, > never escalates to error), W1227 cmd_orphan_routes (36th, per-route > dead-endpoint projection; single closed-enum rule orphan-route with > confidence-banded per-result level: high + medium → warning, low → > note; warning ceiling — heuristic detector, never escalates to error; > the used bucket is filtered upstream so SARIF consumers never see > non-actionable rows), W1229 cmd_verify_imports (37th, first SHIP > emitter that escalates to error — two closed-enum rules: > invalid-import (warning) for unresolved with FTS5 fuzzy-match > candidates, hallucination-import (error) for unresolved with no > candidates; verify-imports is the canonical "hallucination firewall" > detector for LLM-era code and the only verify-imports rule that > escalates to error per the W1229 scope discipline). (2) Pattern-2 > variant-D family enablement (W1241) — landed the canonical > resolution_disclosure() helper at > src/roam/output/formatter.py:1263 + _RESOLUTION_KINDS frozen > closed-enum (symbol / file / fuzzy / unresolved) + drift-guard > test (tests/test_resolution_disclosure.py). Helper substrate now > live for the W1242/W1243/W1244 Wave-1 adoption sweep across > cmd_impact / cmd_preflight / cmd_diagnose (in flight at > consolidation time). (3) SKIP-disclosure propagation arc continued > through Wave 14 — _KNOWN_MISSING decremented 20 → 17 via the > three W1226/W1227/W1229 SHIP-promote pin-list removals (no new > docstring waves this batch — long-tail of the propagation arc). > (4) Drift-guard remediation pass — W1237 (cmd_risk edge-kind > vocabulary canonicalized to roam.db.edge_kinds) + W1238 > (catalog/detectors.py framework-detector plugin loop migrated from > bare-except to log.warning(...) + continue per W531 fail-loud > discipline; previously-grandfathered _PRE_W662_PENDING entries > dropped to zero in that file). Plus a 884-LOC research memo > (dev/PATTERN-2-EVOLUTION-2026-05-16.md) cataloguing the > <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> > seven-variant Pattern-2 family taxonomy and seven open gaps.

• SARIF SHIP family grew from 34 to 37 emitters (W1226+W1227+W1229). Added cmd_flag_dead (35th SHIP, three closed-enum rules under flag-* namespace — flag-staleness / flag-single-reference / flag-constant-default; staleness-banded per-result level with a warning ceiling — heuristic detector, never escalates to error), cmd_orphan_routes (36th SHIP, per-route dead-endpoint projection; single closed-enum rule orphan-route; confidence-banded per-result level — high + medium → warning, low → note; warning ceiling — heuristic detector, never escalates to error; the used bucket is filtered upstream so SARIF consumers never see non-actionable rows), cmd_verify_imports (37th SHIP, first SHIP emitter that escalates to error — two closed-enum rules: invalid-import (warning) for unresolved with FTS5 fuzzy-match candidates, hallucination-import (error) for unresolved with no candidates; verify-imports is the canonical "hallucination firewall" detector for LLM-era code and the only verify-imports rule that escalates to error per the W1229 scope discipline; resolved rows are filtered upstream so SARIF consumers never see non-actionable rows). All three wrappers hash-stable additive; no persisted finding rows touched. Hash-stability invariant held across all 37 emitters.
• Pattern-2 variant-D family enablement substrate (W1241). Landed the canonical resolution_disclosure() helper at src/roam/output/formatter.py:1263 + a frozen closed-enum _RESOLUTION_KINDS (symbol / file / fuzzy / unresolved) + drift-guard test at tests/test_resolution_disclosure.py. Implements the W324 cmd_annotate template at substrate level — every command that calls find_symbol() with an implicit fallback chain can now surface which tier of the resolver succeeded. The partial_success flag is True for any non-symbol resolution. Enables the W1242/W1243/W1244 Wave-1 adoption sweep across cmd_impact / cmd_preflight / cmd_diagnose (in flight at consolidation time). The W1233 audit (Pattern-2c on 38 sites — Wave 1+2+3 enumeration) + W1246 audit (cmd_trace find_symbol_id non-compliant → W1248 capture) frame the remaining adoption work as a multi-wave propagation arc structurally parallel to the SARIF SHIP / SKIP-DISCLOSURE arc that drove CONSOLIDATEs 4 → 13.
• SKIP-disclosure propagation — pin-list 20 → 17 via SHIP-promotes (no new docstring waves). Each of W1226 + W1227 + W1229 decremented _KNOWN_MISSING in-batch via the tests/test_known_missing_pin_is_current inverse-drift guard — no second-pass stale-pin sweep required, continuing the W1222 hygiene discipline (CONSOLIDATE-11) and the CONSOLIDATE-12 "every SHIP landing decrements _KNOWN_MISSING" pattern. The propagation arc is now ~91% complete from the original 196-file gap (179 commands closed; 196 → 17). The 17 surviving pin-list entries are all the long-tail audit-needed commands flagged by W1175-RESEARCH (cmd_debt + cmd_entry_points + cmd_guard + cmd_map + cmd_metrics + cmd_path_coverage + cmd_patterns + cmd_plan_refactor + cmd_pytest_fixtures + cmd_risk + cmd_safe_delete + cmd_safe_zones + cmd_simulate_departure + cmd_suggest_refactoring + cmd_testmap + cmd_why_slow + cmd_ws).
• Drift-guard remediation (W1237 + W1238). W1237 canonicalized the edge-kind vocabulary in cmd_risk onto the roam.db.edge_kinds registry — closes a quiet drift class where edge-kind literals could diverge across cmd_*.py callsites. W1238 migrated the framework-detector plugin loop in catalog/detectors.py (previously at lines 2044/2048, drifted to 2153/2157) from bare except Exception: continue|pass to log.warning(...) + continue per W531 fail-loud discipline; the plugin-isolation perimeter rationale is preserved in the inline comments at the call site; the swallow is now visible. Two previously-grandfathered _PRE_W662_PENDING entries dropped to zero in catalog/detectors.py (stale-pin hygiene applied alongside the migration).
• Research memo — dev/PATTERN-2-EVOLUTION-2026-05-16.md (884 LOC). Catalogues the seven-variant Pattern-2 family taxonomy as of 2026-05-16: 2a compound-recipe / 2b empty-corpus / 2c resolution-state (W1241 helper substrate) / 2d producer-gap (W261 redaction reason) / 2e shared-substrate (W1018 YAML loader) / 2f packet-layer (W210 evidence_stale) / 2g closed-vocabulary-unknown (W1077/W1080 structured_unknown). Companion to dev/PATTERN-2-EVOLUTION-2026-05-15.md (yesterday's Python-3.11+ language-feature survey of warnings_out — verdict STAY). Surfaces seven open gaps + counts adoption per variant. No source changes; sequencing decisions are Cranot's.

<!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->

Audits / verdicts — W1226-W1248 batch

• W1230 audit — cmd_test_gaps re-verification — VERDICT SKIP CONFIRMED. Carry-forward audit from W1202 (CONSOLIDATE-9) re-opened by the Wave 14b docstring landing. Re-verified: REPORT-not-detector pattern (no per-location FindingRecord persistence). W1202 close confirmed; SKIP-disclosure docstring stays in place.
• W1231 audit — drift-guard triage (5 failures → 2 fixes + 2 follow-ups). Five drift-guard failures triaged: W1237 (cmd_risk edge_kinds canonicalize — FIXED), W1238 (catalog/detectors.py bare-except migration — FIXED), W1239 (drift-guard hygiene follow-up — FOLLOW-UP), W1240 (drift-guard hygiene follow-up — FOLLOW-UP). Per W531 fail-loud discipline both fixes ship the swallow as a visible log.warning(...) rather than a bare continue|pass.
• W1233 audit — Pattern-2c on 38 sites, Wave 1+2+3 enumeration. Surveyed 38 call sites of find_symbol() / find_symbol_id() across src/roam/commands/. Wave 1 = cmd_impact + cmd_preflight + cmd_diagnose (captured as W1242 + W1243 + W1244 — in flight at consolidation time). Wave 2 = the next 10 sites (captured as W1245). Wave 3 = remaining (captured as W1246 + W1248). Adoption sequencing parallels the SARIF SHIP / SKIP-DISCLOSURE arc — Wave 1 lands the helper substrate consumers, Wave 2/3 sweep the long tail.
• **W1236 audit — cmd_*.py SARIF helper convention sweep — VERDICT BENIGN.** Surveyed in-module SARIF helper definitions across SHIP emitters. Module-local SARIF convention is consistent across all 37 emitters; no canonicalization needed at substrate level. W1247 captured the doc-pass to ensure CLAUDE.md reflects the module-local SARIF convention.
• W1246 audit — cmd_trace.find_symbol_id Pattern-2c — VERDICT NON-COMPLIANT. The find_symbol_id callsite in cmd_trace does not currently emit resolution disclosure on degraded resolution. Captured as W1248 for Wave-3 adoption sweep.
• W1224-impl Wave 14 hygiene discipline held across the post-CONSOLIDATE-12 batch. Each of W1226 + W1227 + W1229 decremented _KNOWN_MISSING in-batch — no stale-pin removal sweep needed, continuing the W1222 carry-forward discipline (CONSOLIDATE-11) and the CONSOLIDATE-12 0-stale-pin invariant.

In flight — W1226-W1248 batch (parallel dispatches not yet on disk)

• W1242 — cmd_impact Pattern-2c adoption + W1243 — cmd_preflight Pattern-2c adoption + W1244 — cmd_diagnose Pattern-2c adoption — dispatched in parallel as the W1241 helper substrate consumers (Wave 1). None touch CHANGELOG / HANDOVER / BACKLOG / SESSION-SNAPSHOT — zero collision risk with this consolidation pass.
• W1239 — drift-guard hygiene follow-up audit dispatched in parallel; non-CHANGELOG-touching.

Added — W1207-W1224 batch (post-CONSOLIDATE-11, 2026-05-16 /loop iteration N+12)

> ~13 completions since CONSOLIDATE-11 (Section 57). Two milestones: > (1) SARIF SHIP family closed out the entire CONSOLIDATE-11 SHIP > candidate roster — 6 emitters landed (W1207 + W1209 + W1210 + W1211 + > W1213 + W1216), bringing the SHIP family from 28 to 34 emitters in a > single window. cmd_llm_smells (29th SHIP, 10 closed-enum rules > under the llm-smells/ namespace — first SHIP emitter with > double-digit closed-enum rule count + severity-banded level) + > cmd_fan (30th SHIP, per-symbol fan-in/fan-out projection) + > cmd_hotspots (31st SHIP, runtime-mode only; --security/--danger > sub-modes emit raw findings outside the closed-enum > hotspots/* rule catalogue per W1210's scope discipline) + > cmd_dark_matter (32nd SHIP, per-pair hidden-coupling projection; > single closed-enum rule dark-matter/hidden-coupling with > confidence-tier-banded severity) + cmd_duplicates (33rd SHIP, BAIL- > and-capture promotion landing — per-cluster semantic-duplicate > projection; single closed-enum rule duplicates/cluster with > similarity-banded severity) + cmd_laws (34th SHIP, per-rule > invariant projection from the W119 mined-laws substrate). All six > wrappers hash-stable additive; no persisted finding rows touched. > The 6-candidate SHIP roster from CONSOLIDATE-11 closed to zero > outstanding in this window. (2) Pattern-3b propagation arc — > 14+ waves shipped, _KNOWN_MISSING 64 → 20. W1224-impl landed > 37 SKIP-eligible docstrings across two waves (Wave 14a = 15 > docstrings; Wave 14b = 22 docstrings) — the largest single-wave > batch of the arc to date, and 0 BAILs across both sub-waves. > Sites: cut / dev_profile / doc_staleness / docs_coverage / drift / > effects / eval_retrieve / evidence_diff / evidence_doctor / fitness / > fn_coupling / graph_stats / idempotency / index / index_bundle (14a) + > ingest_trace / invariants / mutate / owner / pr_diff / pr_prep / > side_effects / split / stats / suggest_reviewers / surface / > syntax_check / telemetry / test_gaps / test_pyramid / tx_boundaries / > version / vuln_map / vuln_reach / workflow / xlang / index_stats (14b). > Pin-list now down to 20 surviving entries (cmd_debt + > cmd_entry_points + cmd_flag_dead + cmd_guard + cmd_map + > cmd_metrics + cmd_orphan_routes + cmd_path_coverage + > cmd_patterns + cmd_plan_refactor + cmd_pytest_fixtures + > cmd_risk + cmd_safe_delete + cmd_safe_zones + > cmd_simulate_departure + cmd_suggest_refactoring + > cmd_testmap + cmd_verify_imports + cmd_why_slow + cmd_ws).

• SARIF SHIP family grew from 28 to 34 emitters (W1207+W1209+W1210+W1211+W1213+W1216). Added cmd_llm_smells (29th SHIP, 10 closed-enum rules under the llm-smells/ namespace — severity-banded per-result level; per-occurrence LLM-API anti-pattern projection; first SHIP emitter with double-digit closed-enum rule count), cmd_fan (30th SHIP, per-symbol fan-in/fan-out projection via fan_to_sarif), cmd_hotspots (31st SHIP, runtime-mode only via hotspots_to_sarif; --security and --danger sub-modes emit raw findings outside the closed-enum hotspots/* rule catalogue per W1210's scope discipline — first SHIP emitter with mode-conditional rule-catalogue scoping), cmd_dark_matter (32nd SHIP, per-pair hidden-coupling projection via dark_matter_to_sarif; single closed-enum rule dark-matter/hidden-coupling with confidence-tier-banded severity), cmd_duplicates (33rd SHIP, BAIL-and-capture promotion landing — per-cluster semantic-duplicate projection via duplicates_to_sarif; single closed-enum rule duplicates/cluster with similarity-banded severity), cmd_laws (34th SHIP, per-rule mined-invariant projection via laws_to_sarif). All six wrappers hash-stable additive; no persisted finding rows touched.
• Pattern-3b propagation arc — 14+ waves shipped, _KNOWN_MISSING 64 → 20. Wave 14 (W1224-impl Wave 14a + Wave 14b) landed 37 SKIP-eligible docstrings across two contiguous sub-waves with 0 BAILs each. Wave 14a (15 docstrings): cut / dev-profile / doc-staleness / docs-coverage / drift / effects / eval-retrieve / evidence-diff / evidence-doctor / fitness / fn-coupling / graph-stats / idempotency / index / index-bundle. Wave 14b (22 docstrings): ingest-trace / invariants / mutate / owner / pr-diff / pr-prep / side-effects / split / stats / suggest-reviewers / surface / syntax-check / telemetry / test-gaps / test-pyramid / tx-boundaries / version / vuln-map / vuln-reach / workflow / xlang / index-stats. Largest single-wave batch of the arc to date — 37 docstrings in one window vs the prior maximum of 12 (W1187-impl Wave 4) and 11 (W1188-impl Wave 5 / W1191-impl Wave 7). Each Wave 14 site BAIL-checked clean (no emit_finding / findings_store.persist call site at the destination).
• 6-candidate SHIP roster from CONSOLIDATE-11 closed to zero outstanding. All six SHIP candidates that CONSOLIDATE-11 deferred (W1207 / W1209 / W1210 / W1211 / W1213 / W1216) shipped in this window. The capture-then-defer discipline (BAIL-and-capture pattern from CONSOLIDATE-10 W1206-impl-skip; deeper-audit reclassification pattern from CONSOLIDATE-10 W1206-audit-unclear) achieved its terminal state — every captured SHIP candidate followed by a clean SHIP landing on next dispatch.

Audits / verdicts — W1207-W1224 batch

• W1224-impl (Wave 14a + Wave 14b) — 37 SKIP-eligible docstring sites, 0 BAILs across both sub-waves. Wave 14a covered the invocation-scoped-aggregate + state-mutating + validator slice (cut / dev-profile / doc-staleness / docs-coverage / drift / effects / eval-retrieve / evidence-diff / evidence-doctor / fitness / fn-coupling / graph-stats / idempotency / index / index-bundle). Wave 14b covered the aggregate / composer / state-mutating / validator slice (ingest-trace / invariants / mutate / owner / pr-diff / pr-prep / side-effects / split / stats / suggest-reviewers / surface / syntax-check / telemetry / test-gaps / test-pyramid / tx-boundaries / version / vuln-map / vuln-reach / workflow / xlang / index-stats). Premise checks uniform on both sub-waves — no per-location FindingRecord persistence; no BAIL discoveries; no reclassifications. Largest single-wave batch of the arc to date.
• W1222 pin-list hygiene discipline codified (drive-by carry-forward). The "every SHIP landing decrements _KNOWN_MISSING" discipline surfaced by W1222 (CONSOLIDATE-11 — cmd_over_fetch was still pinned despite W1219 having landed) held across all 6 SHIP landings in this window. Each of the 6 emitter landings (W1207 / W1209 / W1210 / W1211 / W1213 / W1216) updated _KNOWN_MISSING in-batch via the tests/test_known_missing_pin_is_current inverse-drift guard — no second-pass stale-pin sweep required.

In flight — W1207-W1224 batch (parallel dispatches not yet on disk)

• W1210 + W1211 — cmd_hotspots + cmd_dark_matter SHIP impl were dispatched in parallel to the docstring waves; both have landed their SHIP wrappers + SARIF projection in _SARIF_CONSUMERS (visible in the cli.py 34-entry tuple). No additional concurrent waves expected to land on CHANGELOG / HANDOVER / BACKLOG / SESSION-SNAPSHOT in this window.

Added — W1213-W1222 batch (post-CONSOLIDATE-10, 2026-05-16 /loop iteration N+11)

> ~10 completions since CONSOLIDATE-10 (Section 56). Two milestones: > (1) SARIF SHIP family grew from 24 to 28 emitters (W1208 + W1217 + > W1218 + W1219 + W1215). cmd_n1 (24th, W110 N+1 detector wrapper > with 3 closed-enum rules — high/med/low; 89+30 tests pass) + > cmd_missing_index (25th, 3 closed-enum rules; 20 tests pass) + > cmd_orphan_imports (26th, 3 closed-enum rules — > internal_typo=error / missing_package=warning / > missing_local=warning) + cmd_over_fetch (27th, single closed-enum > rule at warning; dual-shape endpoint+model handling) + > cmd_bus_factor (28th, 3 closed-enum rules — concentration / > stale-ownership / solo-summary; directory-anchor pattern; > hash-stable sha256 verified). (2) Pattern-3b propagation arc — > 12+ waves shipped, _KNOWN_MISSING 96 → 64. Wave 13 closed 10 > SKIP-eligible docstrings with zero BAILs (changelog / db_check / > intent_check / metrics_push / recommend / report / retrieve / > schema / search_semantic / simulate). W1212 reclassification + > W1220 SKIP + W1222 inline stale-pin removal close this batch's > propagation contribution. 6 SHIP candidates remain pending (W1207 > cmd_llm_smells / W1209 cmd_fan / W1210 cmd_hotspots / > W1211 cmd_dark_matter / W1213 cmd_duplicates / W1216 cmd_laws).

• SARIF SHIP family grew to 28 emitters (W1208+W1217+W1218+W1219+W1215). Added cmd_n1 (24th, W110 N+1 detector with 3 closed-enum rules; 89+30 tests pass), cmd_missing_index (25th, 3 closed-enum rules; 20 tests pass), cmd_orphan_imports (26th, 3 confidence tiers — internal_typo=error / missing_package=warning / missing_local=warning), cmd_over_fetch (27th, single closed-enum rule, dual endpoint+model handling), cmd_bus_factor (28th, directory-anchor pattern with 3 rules — concentration / stale-ownership / solo-summary; hash-stable sha256 verified). All five additive wrappers; hash-stability invariant held across all 28 emitters.
• Pattern-3b propagation arc — 12+ waves shipped, _KNOWN_MISSING 96→64. Wave 13 (W1221-audit + W1221-impl) landed 10 SKIP-eligible docstrings with 0 BAILs (changelog / db_check / intent_check / metrics_push / recommend / report / retrieve / schema / search_semantic / simulate). W1212 reclassification (REPORT-not-detector for cmd_coverage_gaps) + W1220 SKIP (cmd_capabilities capability-registry manifest emitter) + W1222 inline (stale-pin removal from _KNOWN_MISSING per W1219 follow-up) close this batch's propagation contribution. ~22 commands closed across propagation + SHIP-promote in this batch.
• W1213 — cmd_duplicates SHIP CAPTURED (BAIL discovery in W1206-impl-skip). Captured cleanly via the BAIL-and-capture pattern; pending impl as SHIP candidate per the W1207-W1213 + W1216 roster.

Audits / verdicts — W1213-W1222 batch

• W1221-audit (Wave 13) — 10 SKIP-eligible docstring sites, 0 BAILs. Sites: cmd_changelog + cmd_db_check + cmd_intent_check + cmd_metrics_push + cmd_recommend + cmd_report + cmd_retrieve + cmd_schema + cmd_search_semantic + cmd_simulate. Premise checks passed uniformly — no per-location FindingRecord persistence; no BAIL discoveries; no reclassifications.
• W1222 — cmd_over_fetch stale-pin removal from _KNOWN_MISSING. Inline follow-up of W1219 SHIP — the over_fetch entry was still pinned in the disclosure-coverage _KNOWN_MISSING list despite the SARIF wrapper having landed; removed in-batch.

Added — W1199-W1212 batch (post-CONSOLIDATE-9, 2026-05-16 /loop iteration N+10)

> ~15 completions since CONSOLIDATE-9 (Section 55). Three milestones > + one reclassification discipline: (1) SARIF SHIP family grew to > 23-24 emitters (W1203 + W1208). cmd_test_impact (23rd SHIP, ~333 > LOC = 160 prod + 173 test) joined as a per-test reach_count ranker > with file-level anchor, reusing the global --sarif flag plumbed > through _SARIF_CONSUMERS; 11 new SARIF tests + 59 pre-existing > pass. cmd_n1 (24th SHIP) joined as a W110 N+1 detector SARIF > wrapper with per-query findings. (2) Pattern-3b propagation arc — > 11 waves shipped, 58% gap closed. _KNOWN_MISSING dropped 96 → 82 > across Wave 10 (W1205-impl, 10 Bucket B docstrings; 96 → 86) and > Wave 11 (W1206-impl-skip, 5 of 6 SKIP docstrings; 88 → 82). The > 6th SKIP docstring (cmd_duplicates) bailed mid-impl and was > captured as W1213 SHIP. 114 commands closed across W1180 → W1212 > (58% of the original 196-file gap). (3) Reclassification > discipline — W1212 + W1213. W1199 (CONSOLIDATE-9 SHIP candidate > for cmd_coverage_gaps) was REVISED to SKIP-DISCLOSURE this window > via W1206-audit-unclear's deeper audit (REPORT command — wrap_findings > is envelope-level, not per-location). Symmetrically, cmd_duplicates > was discovered as a SHIP candidate by W1206-impl-skip's premise check > and captured as W1213. The methodological move is deeper audit > beats initial classification.

• SARIF SHIP family at 23-24 emitters (W1203 + W1208). cmd_test_impact (per-test reach_count ranker, file-level anchor; reuses global --sarif flag via _SARIF_CONSUMERS; ~333 LOC = 160 prod + 173 test; 11 new SARIF tests + 59 pre-existing pass; hash-stable additive wrapper) + cmd_n1 (W110 N+1 detector wrapper with per-query findings; hash-stable additive wrapper).
• Pattern-3b propagation arc — 11 waves shipped, 58% gap closed. _KNOWN_MISSING 96 → 82 across W1205-impl + W1206-impl-skip (114 commands closed across W1180 → W1212). Wave 10 added 10 Bucket B docstrings (cmd_batch_search + cmd_file + cmd_symbol + cmd_relate + cmd_refs_text + cmd_history_grep + cmd_recipes + cmd_sketch + cmd_pr_analyze + cmd_pr_replay); Wave 11 added 5 SKIP docstrings (cmd_affected + cmd_closure + cmd_compare + cmd_conventions + cmd_causal_graph) with cmd_duplicates bailed and captured as W1213.
• Reclassification discipline applied (W1212 + W1213). cmd_coverage_gaps reclassified from W1199-SHIP (CONSOLIDATE-9) to W1212-SKIP-DISCLOSURE (REPORT command, not per-location detector — wrap_findings is envelope-level, no FindingRecord persistence). cmd_duplicates discovered as SHIP via BAIL-and-capture from W1206-impl-skip premise failure. First formal cross-session reclassification in the propagation arc; symmetric direction (SHIP→SKIP and SKIP→SHIP) demonstrated.

Audits / verdicts — W1199-W1212 batch

• W1205-audit — Wave 10: 10 Bucket B docstring sites — VERDICT SKIP-DISCLOSURE x10. All 10 docstring landings shipped via W1205-impl. _KNOWN_MISSING 96 → 86.
• W1206-audit — Wave 11 mixed batch: 6 SKIP + 4 unclear + 2 SHIP. SHIP candidates captured as W1207 (cmd_llm_smells) + W1208 (cmd_n1, shipped this window). 4 unclear resolved via W1206-audit-unclear (below).
• W1206-audit-unclear — 4-command deeper audit. 3 SHIP captured (cmd_fan W1209 / cmd_hotspots W1210 / cmd_dark_matter W1211; ~1-2d each); 1 REVISED SKIP (cmd_coverage_gaps is REPORT-not-detector → W1212 supersedes W1199). The deeper-audit reclassification is the first formal cross-session classification revision in the Pattern-3b propagation arc.
• W1212 — cmd_coverage_gaps REVISED SKIP-DISCLOSURE. Supersedes W1199 SHIP from CONSOLIDATE-9. REPORT command — wrap_findings stays envelope-level; no FindingRecord. ~10 LOC docstring.

Added — W1186-W1198 batch (post-CONSOLIDATE-8, 2026-05-16 /loop iteration N+9)

> ~25 completions since CONSOLIDATE-8 (Section 54). Three milestones: > (1) SARIF SHIP family at 22 emitters (W1192 + W1195). cmd_delete_check > (21st SHIP, ~165 LOC) joined as the first SHIP emitter with PRIMARY + > SECONDARY SARIF locations (deletion candidate is PRIMARY; surviving > refs in code/test/docs/config are SECONDARY) for the BREAK-RISK > gate-blocking pattern. cmd_auth_gaps (22nd SHIP, ~180 LOC) joined as > the first SHIP emitter with explicit 3-tier confidence in SARIF output > (static_analysis / structural / heuristic flow from > single-source-of-truth confidence map into properties.confidence). > Pre-batch the SHIP family was 20 (cmd_smells + cmd_clones + > cmd_partition + cmd_affected_tests + cmd_impact + cmd_critique + 14 > pre-existing). (2) Pattern-3b propagation arc — 9 waves shipped, > 51% gap closed. _KNOWN_MISSING dropped 196 → 96 across W1180 + > W1181 + W1182 + W1185 + W1187 + W1188 + W1189 + W1190 + W1191 + W1194 > + W1195 + W1197 + W1198 (100 commands closed; 51% of the original > 196-file gap). This batch added 4 more waves on top of Section 54's 3: > Wave 6 (W1189-impl, 10 commands; 137 → 127), Wave 7 (W1191-impl, 11 > commands + cmd_delete_check stale-pin removal drive-by; 125 → 114), > Wave 8 (W1194-impl, 10 Bucket B/C/E; 113 → 103), Wave 9 (W1197-impl, > 4 SKIP + 2 UNCLEAR-resolved-to-SKIP; 100 → 96). (3) Capture > discipline preserved — 6 SHIP candidates deferred cleanly > (W1199-W1204). W1198-audit identified cmd_coverage_gaps (W1199) > + cmd_orphan_routes (W1200) + cmd_pytest_fixtures (W1201) + > cmd_test_gaps (W1202) + cmd_test_impact (W1203) + > cmd_verify_imports (W1204); ~7-10d total effort. All 6 emit > per-location findings in JSON envelope today; remaining work is > emit_finding() integration + SARIF wrapper per W1192/W1195 scaffold. > Hash-stability invariant held across all 22 emitters. 131/131 SARIF > tests pass throughout.

• SARIF SHIP family at 22 emitters (W1192 + W1195). cmd_delete_check (BREAK-RISK gate-blocking, PRIMARY+SECONDARY locations — first SHIP emitter with multi-location pattern; ~165 LOC; hash-stable; 131/131 SARIF tests pass) + cmd_auth_gaps (3-tier confidence: static_analysis/structural/heuristic, reuses single-source-of-truth confidence mapping; first SHIP emitter with explicit 3-tier confidence in SARIF output; ~180 LOC; hash-stable; 131/131 SARIF tests pass).
• Pattern-3b propagation arc — 9 waves shipped, 51% gap closed. _KNOWN_MISSING 196 → 96 across W1180+W1181+W1182+W1185+W1187+W1188+W1189+W1190+W1191+W1194+W1195+W1197+W1198. Per-wave throughput: 10-12 docstrings; audit-and-emit asymmetric pattern; cryptographic hash-stability where required. Section 55 ships Wave 6 (W1189-impl), Wave 7 (W1191-impl), Wave 8 (W1194-impl), Wave 9 (W1197-impl) on top of Section 54's Waves 3-5.
• SHIP candidate pipeline captured (W1199-W1204). 6 commands deferred from W1198-audit: cmd_coverage_gaps / cmd_orphan_routes / cmd_pytest_fixtures / cmd_test_gaps / cmd_test_impact / cmd_verify_imports. ~7-10d total effort (1-2d each). All emit per-location findings in JSON envelope today; need emit_finding() integration + SARIF wrapper per W1192/W1195 scaffold. Capture discipline preserved — surfaced inside next-wave batch and deferred cleanly with file paths + effort estimates rather than collapsing into the same window.

Audits / verdicts — W1186-W1198 batch

• W1192-audit — cmd_delete_check SHIP (21st SARIF emitter) + cmd_migration_safety SKIP (validator-not-detector). First SHIP emitter with PRIMARY + SECONDARY SARIF locations; pattern portable to any "X is referenced by Y[]" gate-blocking check.
• W1195-audit — cmd_auth_gaps SHIP (22nd SARIF emitter) + cmd_audit_trail_verify SKIP (verifier-not-detector). First SHIP emitter with explicit 3-tier confidence (static_analysis/structural/heuristic) in SARIF output; pattern portable to any emitter publishing confidence tiers in its JSON envelope.
• W1188-audit / W1189-audit / W1191-audit / W1194-audit / W1197-audit — SKIP-DISCLOSURE x52 + 6 SHIP candidates + 2 UNCLEAR→SKIP. Five propagation-wave verdicts spanning Waves 5-9. The Wave 7 audit surfaced cmd_delete_check as a stale-pin candidate; promoted to SHIP via W1192. The Wave 9 audit surfaced 6 SHIP candidates (W1199-W1204) deferred with effort estimates.
• W1198-audit — 6 SHIP candidates (W1199-W1204) + 2 UNCLEAR → SKIP. Capture-over-implementation discipline preserved. SHIP candidates: cmd_coverage_gaps, cmd_orphan_routes, cmd_pytest_fixtures, cmd_test_gaps, cmd_test_impact, cmd_verify_imports. ~7-10d total effort across the six.
• W1196 — breaking_to_sarif() dormant code investigation. CAPTURED only; no code change. Captured for future close-out wave.

Added — W1186-W1189 batch (post-CONSOLIDATE-7, 2026-05-16 /loop iteration N+8)

> ~10 completions since CONSOLIDATE-7 (Section 53). Three pillars: > (1) SARIF substrate adoption STRUCTURALLY COMPLETE — all 19 > *_to_sarif helpers across the codebase now use _rule_entry() + > _result_entry() factories (W1178 + W1179a + W1179b + W1186 polish). > Hash-stability cryptographically verified via sha256 matches on > pre/post adopter outputs. Net ~LOC-neutral overall (per W1080 > discipline) — the substrate's value is structural API consistency, > not raw LOC reduction. (2) Pattern-3b propagation arc — 5 waves > shipped: Wave 1 (10 bootstrap) + Wave 2 (10 local-state) + W1185 > outliers (2 commands) + Wave 3 (12 codegen) + Wave 4 (12 > exploration) + Wave 5 (11 continuation) = 56 commands closed. > _KNOWN_MISSING dropped 196 → 138 across this stretch (29% of the > original gap closed). Wave 6 audit landed via W1189-audit (10 > commands queued for next dispatch). (3) Concurrent-merge > discipline battle-tested — multiple "file modified since read" > guards fired across W1179a/b + W1180/W1181/W1185 races and > resolved cleanly via the Edit guard's read-before-write contract.

• SARIF substrate adoption STRUCTURALLY COMPLETE (W1179a + W1179b + W1186). All 19 *_to_sarif helpers across the codebase use _rule_entry() + _result_entry() factories from src/roam/output/sarif.py. W1179a shipped 8 emitter substrate adoption with hash-stability cryptographically verified (sha256 matches on pre/post SARIF outputs); ~LOC-neutral (honest discipline per W1080). W1179b shipped 8 more emitters — PARTIAL extraction is now STRUCTURALLY COMPLETE across all 19 emitters. W1186 polished the substrate by adding an extras parameter to _rule_entry() (mirroring the _result_entry() extras pattern) + 1 inline refactor (taint emitter). 131/131 SARIF tests pass.
• Pattern-3b propagation arc — 5 waves shipped (W1180 + W1181 + W1185 + W1182 + W1187 + W1188). _KNOWN_MISSING dropped 196 → 138 across this stretch (56 commands closed). Wave 1 (W1180): 10 bootstrap commands. Wave 2 (W1181): 10 local-state commands. W1185 outliers: 2 commands. Wave 3 (W1182-impl, +12 Bucket C codegen docstrings): _KNOWN_MISSING 173 → 161. Drive-by: cmd_lsp anchor newline fix. Wave 4 (W1187-audit + W1187-impl, +12 Bucket B exploration/aggregate docstrings): _KNOWN_MISSING 161 → 149. Wave 5 (W1188-audit + W1188-impl, +11 Bucket B continuation docstrings): _KNOWN_MISSING 149 → 138. Wave 6 audit landed (W1189-audit) — 10 candidates queued: cmd_help_search + cmd_timeline + cmd_trends + cmd_alerts + cmd_weather + cmd_ai_ratio + cmd_ai_readiness + cmd_dogfood + cmd_postmortem + cmd_dogfood_aggregate.
• Concurrent-merge discipline battle-tested. Multiple src/roam/output/sarif.py + src/roam/cli.py + tests/test_sarif_disclosure_coverage.py edits raced across 5+ parallel waves (W1179a + W1179b + W1180 + W1181 + W1185 + W1182-impl + W1187-impl + W1188-impl). The harness's file-read-before-write Edit guard surfaced the "file modified since read" conflict on every race and forced re-reading; every wave merged cleanly. Third concurrent-merge dance documented this session (preceded by W1159+W1160 in Section 52 and W1180+W1181-impl+W1185-impl in Section 53).

Audits / verdicts — W1186-W1189 batch

• W1187-audit — 12 Bucket B exploration/aggregate commands — VERDICT SKIP-DISCLOSURE x12. Wave 4 batch. Sibling pattern to W1182-audit (codegen-not-analysis) and W1148 (aggregate-not-located-finding). All 12 docstring landings shipped via W1187-impl.
• W1188-audit — 11 Bucket B continuation commands — VERDICT SKIP-DISCLOSURE x11. Wave 5 batch. Continuation of Bucket B exploration/aggregate pattern; all 11 docstring landings shipped via W1188-impl. _KNOWN_MISSING 149 → 138.
• W1189-audit — Wave 6 batch identified (10 commands). cmd_help_search + cmd_timeline + cmd_trends + cmd_alerts + cmd_weather + cmd_ai_ratio + cmd_ai_readiness + cmd_dogfood + cmd_postmortem + cmd_dogfood_aggregate. Queued for next dispatch (W1190+ impl).

Added — W1177-W1185 batch (post-CONSOLIDATE-6, 2026-05-16 /loop iteration N+7)

> ~14 completions since CONSOLIDATE-6 (Section 52) including 8 SHIPPED > outcomes + 2 RESEARCH memos + 3 AUDIT-VERDICTs + 1 PARTIAL audit-extraction. > Three major systemic shifts landed: (1) the SARIF helper substrate > launched via W1178 — _rule_entry() + _result_entry() factories in > sarif.py reduce ~80 LOC of new substrate + ~50 LOC of subtractive > adoption across cmd_dead + cmd_critique + cmd_partition (3 adopters); > 17 more emitters being refactored in parallel (W1179a/b in flight); (2) the > Pattern-3b SARIF-disclosure propagation arc launched via W1175-RESEARCH — > 684-line memo planned 30-50 batches with asymmetric propagation (bulk for > ~135 likely-SKIP, 1:1 for ~14-20 likely-SHIP, ~17 unclear). Wave 1 (W1180, > 10 bootstrap commands) + Wave 2 (W1181-impl, 10 local-state commands) + > W1185 outliers (cmd_lsp + cmd_rules_validate) shipped — _KNOWN_MISSING > dropped from 196 to 174 (33 done, 162 to go); (3) vocabulary canonicalization > disciplines W1156 + W1162 + W1176 sealed with cmd_pr_analyze NO_CHANGES → > NOCHANGES sweep landing this batch and the REFERENCE_REMOVAL_VERDICTS > substrate fully operational. Hash-stability invariant held across all > shipped impls.

• SARIF helper substrate launched (W1177-audit + W1178 + W1179a/b in flight). W1177-audit extracted PARTIAL verdict from the 5-phase pipeline survey across 20 SARIF emitters: 3 patterns identified (Fixed-rule, Dynamic-rule, Complex-multi-rule); ~500 LOC subtractive ceiling. W1178 shipped _rule_entry() + _result_entry() factories in src/roam/output/sarif.py (~80 LOC helpers) + 3 adopters (cmd_dead + cmd_critique + cmd_partition) (~50 LOC subtractive). 131/131 SARIF tests pass. W1179a + W1179b refactoring 17 more emitters in parallel; close-out captured at the next CONSOLIDATE checkpoint.
• Pattern-3b propagation arc launched — Wave 1 + Wave 2 + W1185 outliers (W1175-RESEARCH + W1180 + W1181-audit + W1181-impl + W1185-audit + W1185-impl). <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> W1175-RESEARCH (684-line memo at dev/SARIF-DISCLOSURE-196-PROPAGATION-PLAN-2026-05-16.md) inventoried the 196 unaudited cmd_*.py files: ~135 likely-SKIP + ~14-20 likely-SHIP + ~17 unclear. Asymmetric propagation pattern adopted: bulk audit-and-emit for SKIP; 1:1 audit-then-impl for SHIP. Wave 1 (W1180, +95 LOC across 10 bootstrap commands): SARIF-skip docstring rationale propagated; drive-by pruned 10 stale _KNOWN_MISSING pins. Wave 2 (W1181-audit + W1181-impl, 10 Bucket D commands): substrate-state nouns documented (concurrent merge with W1180/W1185-impl absorbed cleanly). W1185 outliers (W1185-audit + W1185-impl, +15 LOC): cmd_lsp SKIP (editor protocol, not CI/findings) + cmd_rules_validate SKIP (validator-not-detector) docstrings landed. _KNOWN_MISSING dropped from 196 to 174 across the 3 waves combined (33 done, 162 to go). W1182-audit identified 12 Bucket C codegen commands for Wave 3 (cmd_attest + cmd_capsule + cmd_agent_export + cmd_agents_md + cmd_graph_export + cmd_cga + cmd_sbom + cmd_skill_generate + cmd_pr_comment_render + cmd_audit_trail_export + cmd_evidence_oscal + cmd_fingerprint).
• Vocabulary canonicalization disciplines extended (W1151 + W1156 + W1162 + W1164 + W1176). W1151 removed cargo-cult .upper() from _to_level() across 7 sites in sarif.py — hash-stable. W1156 REFERENCE_REMOVAL_VERDICTS frozenset substrate fully operational (~100 LOC; carry-forward from W1156-CONSOLIDATE). W1162 cmd_flag_dead.py canonical "likely_stale" with display "likely-stale" preserved via _STALENESS_DISPLAY map (mirrors W1156 dual-form normalization pattern). W1176 realises the W1164 audit verdict: cmd_pr_analyze NO_CHANGES → NOCHANGES (3 LOC; sibling-aligned bare UPPERCASE).

Audits / verdicts — W1177-W1185 batch

• W1177-audit — SARIF helper-substrate extraction audit — VERDICT PARTIAL EXTRACTION. 5-phase pipeline surveyed across 20 SARIF emitters: 3 patterns identified (Fixed-rule, Dynamic-rule, Complex-multi-rule). _rule_entry() + _result_entry() helpers viable; ~500 LOC subtractive ceiling. Realised via W1178.
• W1181-audit — 10 Bucket D local-state commands — VERDICT SKIP-DISCLOSURE x10. Substrate-state nouns: mode / runs / lease / memory / permits / annotations / suppress / replay / agent-score / agents-md. None are file:line findings emitters; all surface state stored under .roam/. Verdicts landed via W1181-impl docstrings. cmd_lsp + cmd_rules_validate flagged as outliers (handled via W1185-audit + W1185-impl).
• W1182-audit — 12 Bucket C codegen commands — VERDICT pending Wave 3 impl. cmd_attest + cmd_capsule + cmd_agent_export + cmd_agents_md + cmd_graph_export + cmd_cga + cmd_sbom + cmd_skill_generate + cmd_pr_comment_render + cmd_audit_trail_export + cmd_evidence_oscal + cmd_fingerprint. Codegen-artifact-not-analysis rationale (sibling pattern to W1174). Wave 3 impl in flight.
• W1185-audit — 2 outlier commands — VERDICT SKIP x2. cmd_lsp SKIP (editor protocol, not CI); cmd_rules_validate SKIP (validator-not-detector — rules check existing rule definitions, do not analyze code).

Added — W1158-W1176 batch (post-W1156-CONSOLIDATE, 2026-05-16 /loop iteration N+6)

> ~1000+ LOC of impl across 8 SHIPPED outcomes + 2 RESEARCH memos + > 3 AUDIT-VERDICTs. 13 completions since W1156-CONSOLIDATE. Three > structural inflections landed: (1) the SARIF SHIP family grew from > 17 commands (post-W1146) to 20 commands via cmd_impact + > cmd_affected_tests + cmd_partition (smells / clones SHIP impls > in flight as W1171 + W1172); (2) the W1169 SARIF-disclosure-coverage > CI lint discovered 196 unaudited cmd_*.py files, vastly exceeding > the W1166-RESEARCH 4-8 estimate, with a _KNOWN_MISSING frozenset > pinning the gap and W1175-RESEARCH planning the propagation strategy; > (3) two vocabulary canonicalization sweeps closed (W1162 likely-stale > + W1176 NO_CHANGES → NOCHANGES) extending the W1156 dual-form pattern. > action.yml allowlist intent documented (W1167) + cli.py ⊃ action.yml > subset CI lint pinned (W1168).

• SARIF SHIP family expanded to 20 commands (W1159 + W1160 + W1165). cmd_impact SHIP SARIF (~413 LOC across 6 files) emits 4 finding families — affected-file (importance→severity), direct-dependent, sf-convention-test, indirect-ref — and lifts _SARIF_CONSUMERS 15→16. cmd_partition SHIP SARIF (~189 LOC) emits PRIMARY + up-to-10 SECONDARY locations with conflict_risk severity scaling (_SARIF_CONSUMERS 17→18). cmd_affected_tests SHIP SARIF (~147 LOC) uses 3 closed-enum rules (direct=error / transitive=warning / colocated=note) (_SARIF_CONSUMERS 16→17). Concurrent-merge dance between W1159 + W1160 surfaced and resolved cleanly via the Edit guard. cmd_smells SHIP (~250 LOC) + cmd_clones SHIP (~300 LOC, dual-location) in flight as W1171 + W1172. Companion verdicts: cmd_vibe_check SKIP-DISCLOSURE (aggregate; W1170-bundle) + cmd_test_scaffold SKIP-DISCLOSURE (codegen, not analysis; W1170-bundle).
• W332 Pattern-3b SARIF CI-lint substrate complete (W1167 + W1168 + W1169). action.yml _SUPPORTED_SARIF 7-command subset intent documented at the YAML source via a +10 LOC comment block (W1167); cli.py 18-command _SARIF_CONSUMERS ⊃ action.yml 7-command subset relationship pinned via +28 LOC new test_action_yml_supported_sarif_subset_of_cli_consumers in test_sarif_consumer_list.py (W1168); SARIF-disclosure coverage lint (+403 LOC new tests/test_sarif_disclosure_coverage.py) surfaces 196 unaudited cmd_*.py with _KNOWN_MISSING frozenset pin (W1169). Key discovery: the W1166-RESEARCH 4-8 unaudited estimate was off by ~25x; the CI lint baseline pins the gap so future SARIF audits propagate via single-source-of-truth rather than ad-hoc per-command sweeps.
• SARIF-disclosure-pattern maturity memo (W1166-RESEARCH). <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> 555-line memo at dev/SARIF-DISCLOSURE-PATTERN-MATURITY-2026-05-16.md. 14 audits surveyed, ZERO contested verdicts — the SARIF SHIP-vs-SKIP discipline has converged. Top recommendation: ship the W1169 CI lint (DONE this batch). Anti-recommendation: do NOT extract a shared docstring constant — the per-site rationale is load-bearing for future readers anchoring at the call site. Sibling research memo to W1175-RESEARCH (propagation strategy for the 196-file gap; in flight).
• Vocabulary canonicalization sweeps (W1162 + W1176). cmd_flag_dead.py (+13/-5 LOC) canonicalizes "likely-stale" → likely_stale while preserving display form "likely-stale" via new _STALENESS_DISPLAY map — extends the W1156 dual-form normalization pattern. cmd_pr_analyze (3 LOC) renames NO_CHANGES → NOCHANGES (sibling-aligned bare UPPERCASE per W1164 audit verdict option a; 3 sites, no hard-coded test assertions). 154/155 tests pass.
• SARIF-disclosure docstrings on 2 more commands (W1173 + W1174). cmd_vibe_check.py (+8 LOC) names roam findings list --detector vibe-check as the per-finding path; cmd_test_scaffold.py (+8 LOC) anchors codegen-artifact-not-analysis rationale (distinct from W1148's invocation-scoped template). The SKIP-DISCLOSURE docstring pattern now spans 17 commands total when combined with the in-flight W1170-bundle siblings.

Audits / verdicts — W1158-W1176 batch

• W1158 — SARIF source-drift audit — VERDICT HEALTHY DRIFT. 4 SARIF sources surveyed (action.yml 7-cmd subset ⊂ cli.py 18-cmd _SARIF_CONSUMERS; sarif.py 14 emitters + 4 external; 3 cli-only). No bug. Verdict realised via W1167 (action.yml comment) + W1168 (subset lint).
• W1164 — cmd_pr_analyze NO_CHANGES naming audit — VERDICT RENAME to NOCHANGES. Option (a) sibling-aligned bare UPPERCASE chosen over option (b) underscore preservation; 3 sites, no hard-coded test assertions, 5 LOC effort. Realised via W1176.
• W1170-bundle — SARIF audience-disclosure quartet — VERDICT 2x SHIP + 2x SKIP-DISCLOSURE. cmd_smells SHIP (W1171 in flight, ~250 LOC); cmd_clones SHIP (W1172 in flight, ~300 LOC, dual-location pattern); cmd_vibe_check SKIP-DISCLOSURE (aggregate, no file:line; W1173 docstring landed); cmd_test_scaffold SKIP-DISCLOSURE (codegen artifact, not analysis; W1174 docstring landed).

Added — W1149-W1156 batch (post-W1149-CONSOLIDATE, 2026-05-16 /loop iteration N+5)

> ~180 LOC of impl across 5 SHIPPED outcomes + 2 AUDIT-VERDICTs. > 7 completions since W1149-CONSOLIDATE. Two structural verdicts > landed: (1) the SARIF-disclosure pattern now spans 9 commands** > (W1144 + W1145 + W1148 + W1152 + W1154-impl x6), formally documenting > "invocation-scoped aggregates have no SARIF locations[]" as a stable > design rule; (2) reference-removal verdicts (cmd_refs_text + > cmd_delete_check) elevated to a closed-enum frozenset > (REFERENCE_REMOVAL_VERDICTS) via W1156 — drift guard pinned in > test_evidence_v0.py + dual-form normalization preserves CLI display. > publish.yml hardened with persist-credentials:false on build + > smoke checkout steps (W1103) + 3 dist/*.whl sites converted to a > robust single-wheel assertion with quoted variable (W1104).**

• SARIF-skip disclosure pattern formalized across 9 commands (W1154-impl). cmd_orchestrate / cmd_diagnose / cmd_oracle / cmd_plan / cmd_brief / cmd_next now document deliberate SARIF-skip rationale at the module-docstring level (invocation-scoped aggregates; no file:line locations). +36 LOC + 6 blank separators across 6 files. Mirrors the prior W1144 / W1145 / W1148 / W1152 docstrings. Total pattern now spans 9 aggregate-style commands — propagates the "anchor SARIF-skip rationale at the per-site source" discipline so future audits read the call site, not a separate audit memo. W1155 audit pending for the third-tier batch (cmd_fleet / cmd_partition / cmd_affected_tests / cmd_impact / cmd_context).
• Reference-removal verdicts substrate (W1156). New REFERENCE_REMOVAL_VERDICTS frozenset in src/roam/evidence/_vocabulary.py (6 members: safe_to_remove / review / load_bearing / safe / likely_safe / break_risk). Validators wired in cmd_refs_text + cmd_delete_check via _validate_verdict helpers with dual-form normalization — display form "SAFE-TO-REMOVE" round-trips to canonical "safe_to_remove", preserving the CLI human-facing surface while pinning the machine-readable enum. Drift guard pinned in tests/test_evidence_v0.py. 56 + 34 tests pass. ~100 LOC full substrate.
• publish.yml supply-chain hardening (W1103 + W1104). W1103: +6 LOC adds persist-credentials: false to actions/checkout on the build + smoke jobs; publish job correctly untouched (it needs the OIDC token for Trusted Publishing). Removes leaked-credentials attack surface on the non-publish steps. W1104: +36/-3 LOC — 3 dist/*.whl sites (PEP 639 verify + v2-commands verify + SBOM pip install step) converted to a robust single-wheel assertion pattern (shopt -s nullglob + length-1 array check + quoted variable expansion) so the publish workflow fails loudly if dist/ ever contains 0 or 2+ wheels instead of silently degrading on a brace-expansion fallback.

Audits / verdicts — W1149-W1156 batch

• W1154 — SARIF-disclosure audit (6 third-tier commands) — VERDICT SKIP-DISCLOSURE x6. cmd_orchestrate / cmd_diagnose / cmd_oracle / cmd_plan / cmd_brief / cmd_next all classified as aggregate-style commands; none in the SARIF action.yml allowlist. Verdicts landed via W1154-impl docstrings (+36 LOC + 6 blank separators).
• W1134 — reference-removal verdict vocabulary audit — VERDICT LOCAL-CLOSED-ENUM. Reference-removal verdicts (cmd_refs_text + cmd_delete_check) are orthogonal to POLICY_DECISIONS — the former describe whether a code string is safe to delete; the latter describe whether a policy gate passed. Recommendation: stand up a dedicated REFERENCE_REMOVAL_VERDICTS frozenset in evidence/_vocabulary.py rather than overloading POLICY_DECISIONS. Verdict realised by W1156 impl.

Added — W1136-W1149 batch (post-W1133-CONSOLIDATE, 2026-05-15 even more iteration)

> ~410 LOC of impl across 10 SHIPPED outcomes (W1100 +28 LOC + W1099-narrow > ~80 LOC + W1136 +339 LOC + W1141 +36 LOC + W1144 +6 LOC + W1145 +9 LOC + > W1148 +14 LOC) + 1 research memo (W1139-RESEARCH 361 LOC) + 3 SARIF > audit verdicts (W1085 + W1146 + W1147). 11 completions + 4 captures > since W1133-CONSOLIDATE. Three structural outcomes: (a) the W332 > Pattern-3b CLI-boundary thread is functionally closed at v13.x via > the W1141 4th-mirror drift guard; (b) the SARIF audience-disclosure > trilogy (W1144 + W1145 + W1148) propagated the deliberate-skip > rationale docstring across cmd_doctor + cmd_audit + cmd_pr_risk; > (c) the Pattern-3b CLI-arg lint matrix now spans 5 axes (W1111 + > 4 W1121 siblings) PLUS the option-dest extension (W1136 input_path > cluster). The user-facing CLI surface saw the biggest sweep in 30+ > sections: 14 commands gained metavar="SYMBOL" alignment (Strategy D — > no breaking rename) and 6 CLI-only commands gained --file → --path > harmonization with hidden alias backward-compat.

• W332 Pattern-3b CLI-boundary thread functionally closed (W1136 + W1141). New tests/test_w1136_click_option_input_path_dest_lint.py (339 LOC) blocks new @click.option(--input-path) drift at the option-dest axis: 6 canonical sites + 2 legacy carve-outs classified; 4 lint tests + 1 sanity test pass. W1141 (+36 LOC drift guard in tests/test_mcp_param_names.py) pins the _PARAM_ALIASES table for the input_path-cluster as the 4th mirror in the W332 thread (the 3 prior mirrors land at mcp_server.py:_PARAM_ALIASES, test_w1111_click_argument_name_lint.py, test_w1121_click_argument_target_lint.py). The W332 Pattern-3b CLI-boundary thread is now functionally closed at v13.x — per the W1139-RESEARCH coverage matrix, all 6 canonical CLI-side axes are SHIPPED (+ 2 PARTIAL legacy carve-outs documented + 0 GAP).
• SARIF audience-disclosure trilogy (W1144 + W1145 + W1148). Three SARIF skip rationale docstrings landed at the command-module level: W1144 (+6 LOC on cmd_doctor.py:1-12 — environment-scoped diagnostics, no file:line — no SARIF surface), W1145 (+9 LOC on cmd_audit.py — composed-subcommand SARIF flow; no top-level --sarif flag because each subcommand emits its own SARIF if relevant), W1148 (+14/-4 LOC on cmd_pr_risk.py — invocation-scoped aggregates with subject_kind="commit"; action.yml allowlist already excludes pr-risk by design). cmd_critique audited as SHIP candidate (W1146 verdict; W1146-impl in flight). Closes the W1085 → W1146 → W1147 audit triage chain at the documentation layer; the pattern propagates the deliberate-SARIF-skip rationale to the per-site source so future audits anchor at the call site, not at a separate audit memo.
• CLI symbol-cluster metavar alignment (W1100). 14 sites across 11 files got metavar="SYMBOL" (or "SYMBOL_OR_PATH" / "[SYMBOL]" for context-aware variants on cmd_test_scaffold + cmd_testmap) + docstring identifier-tone refresh. ~28 LOC. Hash-stable. cmd_explain_command + cmd_plugins correctly NOT touched (DOMAIN-DISTINCT per W1108/W1120 — CLI command name + plugin name respectively). Strategy D from W1102-RESEARCH: align metavar without a breaking rename, lock the surface via the W1111 + W1121-target AST lints.
• --file → --path harmonization (W1099-narrow). 6 CLI-only commands renamed --file → --path with hidden alias backward-compat: + mcp_server.py + 2 test files updated. ~80 LOC across the cluster. Click's required=True + alias limitation surfaced via cmd_triage (manual UsageError adaptation needed); cmd_pr_bundle deferred to W1141-followup. 508 of 509 tests pass.
• W1139-RESEARCH — Pattern-3b CLI-boundary completeness memo. 361-line memo at dev/PATTERN-3B-CLI-BOUNDARY-COMPLETENESS-2026-05-15.md. Coverage matrix: 6 axes SHIPPED + 2 PARTIAL + 0 GAP. Key finding: W332 functionally closeable in 15 min via the W1141 4th-mirror drift guard (DONE this batch). Companion to W1102-RESEARCH (Section 48) which closed the same question for the @click.argument axis. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->

Audits / verdicts — W1136-W1149 batch

• W1085 — cmd_doctor SARIF audit — VERDICT SKIP-SARIF. Environment-scoped diagnostics (Python version, index status, watcher status, OneDrive/Dropbox detection); no file:line. SARIF surface would be empty. Verdict landed via W1144 docstring (+6 LOC).
• W1146 — cmd_critique SARIF audit — VERDICT SHIP-SARIF. File-located findings: clones-not-edited findings have file:line; impact entries have file paths; intent is diff-wide. Impl dispatched as W1146-impl (in flight as this CONSOLIDATE runs).
• W1147 — cmd_pr_risk SARIF audit — VERDICT SKIP-SARIF. Invocation-scoped aggregates with subject_kind="commit"; SARIF expects file:line subjects. action.yml allowlist already excludes pr-risk by design. Verdict landed via W1148 docstring (+14/-4 LOC).

Added — W1097-W1133 batch (post-W1126-CONSOLIDATE, 2026-05-15 even further iteration)

> ~80 LOC of impl + 253 LOC of new test coverage across 7 SHIPPED > outcomes + 6 audits classified into SYMBOL-CONCEPT or DOMAIN-DISTINCT > verdicts + 1 BAIL (W1101 — premise inverted, captured W1126 inverted > task). 14 completions since W1126-CONSOLIDATE. Two structural > closures: (a) the cmd_runs placeholder-vocabulary cluster > (W1097+W1105+W1116+W1125 = 8 sites swept end-to-end) and (b) the > W1118-bundle reclassification (12 W1111 grandfathered sites > classified into 10 SYMBOL-CONCEPT + 2 DOMAIN-DISTINCT permanent > carve-outs). The v14.0 hard-rename candidate cluster now spans > ~21 files (W1133) — significantly bigger than the W1004 audit's > original 6-file estimate; this scope expansion feeds the W1098 > USER-DECISION at v14.0 planning.

• AST CI lint for @click.argument('target') drift (W1121-target). New tests/test_w1121_click_argument_target_lint.py (253 LOC). 15 sites classified into 4 categories: 13 SYMBOL (joining the v14.0 rename cluster), 1 GIT_REF (bisect start/good/bad), 1 FILE_PATH. Companion lint to W1111 — extends the AST drift-block pattern to a second vocabulary axis. Blocks any 16th drift. Covers the W1099 input_path-cluster gap end-to-end for the target axis.
• W1111 grandfathered-site reclassification (W1106 + W1107 + W1109 + W1118 + W1119 + W1108 + W1120). All 12 W1111 grandfathered sites classified end-to-end: 10 SYMBOL-CONCEPT sites confirmed (cmd_closure.py:191, cmd_testmap.py:232, cmd_impact.py:264, cmd_oracle.py x4 symbol_exists/is_test_only/is_reachable/is_clone_of, cmd_diagnose.py:201, plus the 2 already in the v14.0 cluster) joining the v14.0 hard-rename cluster; 2 DOMAIN-DISTINCT sites (cmd_explain_command.py:150 = CLI command name, cmd_plugins.py:213 = plugin name) marked as permanent grandfather carve-outs. The v14.0 rename cluster now spans ~21 files (W1133 informational capture) versus the original W1004 audit's 6-file scope.
• cmd_runs placeholder-vocabulary structural close (W1125 + W1129 carve-out comments). cmd_runs.py:6 placeholder unified (--action X → <action>); end-to-end closes the cmd_runs placeholder cluster across W1097 + W1105 + W1116 + W1125 (8 sites swept). W1129 (+15 LOC across 3 files) applied the W1108 + W1120 DOMAIN-DISTINCT carve-out comments + the W1111 lint disambiguation comments — anchors the permanent carve-out rationale at the per-site source.
• findings.py vocabulary cross-link cleanup capstone (W1131). +54 LOC across src/roam/db/findings.py: source_version cross-link comment + evidence_json size-GUIDANCE flag + suppressions_json docstring + module-level docstring refresh. Closes the 4-cleanup cluster (W1122 reverse-pointer + W1123 / W1127 / W1128 follow-ups) over the W1126-batch + W1133-batch span. 66/66 findings tests pass; hash-stable (comments + docstring only).
• W1132 — W1111 lint comment update (test-only). tests/test_w1111_click_argument_name_lint.py comment update (~0 LOC, rewording): cmd_impact / cmd_oracle / cmd_diagnose annotations moved from "pending classification" to SYMBOL-CONCEPT confirmed. Net documentation discipline — the W1111 lint's grandfather metadata now reflects the W1118-bundle's classification verdicts.

Audits / verdicts — W1097-W1133 batch

• W1118 + W1119 + W1106 + W1107 + W1109 — VERDICT SYMBOL-CONCEPT (10 sites). All 10 sites resolve <name> to a symbol via find_symbol / find_symbol_with_alternatives — same shape as the v14.0 rename cluster. Optional/default argument shapes preserved at the per-site call.
• W1108 + W1120 — VERDICT DOMAIN-DISTINCT (2 permanent carve-outs). cmd_explain_command.py:150 resolves a CLI command name (not a symbol); cmd_plugins.py:213 resolves a plugin name (not a symbol). Distinct vocabulary axis — out-of-scope for the v14.0 rename. Marked as permanent grandfather; W1129 applied carve-out comments at the per-site source.
• W1133 — INFORMATIONAL capture, v14.0 cluster expansion. The v14.0 hard-rename candidate cluster has grown from the W1004 audit's original 6-file estimate to ~21 files (8 sites on @click.argument("name") + 13 sites on @click.argument("target")). USER DECISION W1098 should reference W1133 for the full v14.0 scope at v14.0 planning.

Added — W1086-W1126 batch (post-W1096-CONSOLIDATE, 2026-05-15 even later iteration)

> ~250 LOC of impl (W1060-take2 + W1086 dominant) + ~120 LOC of new > test coverage + 1 research memo (677 LOC) across 8 shipped outcomes + > 1 BAIL (W1101 premise-inverted). 9 of 9 dispatches closed; 17 new > drive-by W-tasks captured (W1112-W1128). The architectural ship — > to_sarif gained a warnings_out parameter + new closed-enum > descriptor producer.advisory-warning — unlocks 4 sibling SARIF > helpers (W1112-W1115). W1102-RESEARCH closed the W1098 USER-DECISION > as "no v14.0 rename needed; ship the W1111 AST CI lint instead". > Premise-verification-first discipline continues to outperform > force-through (W1101 BAIL: the W1004 audit had misread the dominant > convention, so the proposed sweep would have been backwards).

• SARIF runtime-notification architectural extension (W1060-take2 + W1086). src/roam/output/sarif.py::to_sarif now accepts warnings_out: list[str] + a new closed-enum descriptor producer.advisory-warning on the SARIF tool driver (was missing from the W1046 landing). complexity_to_sarif was wired through with a warnings keyword + src/roam/commands/cmd_complexity.py gained a warnings: list[str] = [] accumulator threaded through 4 envelope sites via the hash-stable omit-when-empty idiom (+84/-29 LOC + 6-test file +39 LOC). 15 SARIF tests pass; hash-stability programmatically asserted for the empty-warnings path. Unblocked W1112-W1115 (4 sibling SARIF helpers: fitness / dead / rules / health) captured as follow-up.
• AST CI lint blocking @click.argument('name') drift (W1111). New tests/test_w1111_click_argument_name_lint.py (199 LOC; 50 LOC executable). 12-file grandfather set (the 6 commands from the W1004 audit + 6 siblings discovered during the lint sweep). Negative path verified — adding a 13th @click.argument("name") site fails the AST scan. Closes the W1098 USER-DECISION via the W1102-RESEARCH deliverable: lock the current drift surface, defer the breaking-change rename until v14.0 ships for unrelated reason.
• roam runs placeholder-vocabulary unification sweep (W1097 + W1105 + W1116). cmd_runs.py placeholders normalised: 1-line NAME → <name> at line 944 (W1097), 7-site --agent NAME → --agent <name> sweep (W1105), 1-line --run-id ID → --run-id <id> (W1116). 14/14 + 4/4 + 29/29 focused tests pass on each leg. All hash-stable comments/help-text only. Drive-by W1117 captured (square-bracket placeholder convention).
• Vocabulary cross-link discipline (W1094 + W1122). +17 LOC across src/roam/evidence/_vocabulary.py + src/roam/output/_severity.py for the severity-vocabulary docstring cross-link (W1094 — closes the W1005 BAIL drive-by). +7-line reverse-pointer comment block in src/roam/db/findings.py:101-107 to evidence.SUBJECT_KINDS (W1122 — closes a drive-by from the W1094 sweep). All 3 sites hash-stable (comments only); 55 W210 drift guards + 92 findings tests pass.
• W1102-RESEARCH — Click-argument rename strategy memo. 677-line research memo at dev/CLICK-ARGUMENT-RENAME-STRATEGY-2026-05-15.md. Key finding: the MCP boundary is already sealed via the _PARAM_ALIASES table landed in W430, so the CLI-side @click.argument("name") drift does NOT silent-fail through MCP. Recommendation: ship the W1111 AST CI lint (DONE) to lock the current 12-site grandfather surface; defer a hard rename until v14.0 ships for an unrelated breaking-change reason. W1098 USER-DECISION downgraded from BLOCKER to FOLLOW-UP. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->

Audits / BAILs — W1086-W1126 batch

• W1101 — memory → memories plural sweep — BAILED (premise inverted). The W1004 audit had read the codebase wrong: ~26 sites use singular-flag/plural-var (the dominant convention) vs ~3 sites with plural-flag. cmd_memory FOLLOWS the dominant convention, so the proposed sweep would have inverted it. New task W1126 — inverted plural-flag harmonize captured to bring the 3 actual outliers in line with the dominant convention.

Added — W1041-W1096 batch (post-W1079-CONSOLIDATE, 2026-05-15 later iteration)

> ~63 LOC + ~39 test LOC across 3 shipped impl waves (W1087/W1091/W1096) + > 8 BAIL/NO-OP/VALIDATED outcomes (W1041/W1004/W1005/W1007/W1008/W1020/W1048/W1060-narrowed) > + 1 helper-confirmed (W1041) — the iteration's load-bearing methodological > output is that BAIL-and-CAPTURE is faster and more accurate than force-through. > ~17 waves across the W1041-W1096 stretch; 4 of 11 dispatches landed > BAIL/NO-OP (W1041 already alphabetical, W1008 already converged via W706+W1057, > W1005 already W547/W564-compliant, W1020 already optimised with scope="module" > overrides where viable); 1 BAIL with prereq capture (W1060-narrowed: warnings_out > accumulator absent in cmd_complexity, so the proposed runtime-notifications > plumb would have been cargo-cult — captured W1084/W1085/W1086 as prereqs > instead); 2 VALIDATED-then-fixed (W1007 → W1091 next_commands LAW 4 fix; > W1008 → drive-by W1093 captured). The bail discipline (W1019b/W1019e/W1080 > precedent + W988+W989 "premise verification is the first step" methodology > from W1001-CONSOLIDATE) generated 9 follow-up W-tasks (W1084-W1097) instead > of forcing-through cargo-cult code. All 11 dispatches used general-purpose > or Explore subagents per the W1072 directive — claude subagent > worktree-MAX_PATH blocker still active on Windows.

• W1087 — CI hardening, 9 jobs got timeout-minutes, 2 workflows got concurrency groups. +17 LOC across .github/workflows/architecture-guardian.yml + .github/workflows/cga-attestation.yml + .github/workflows/roam-ci.yml. architecture-guardian.yml + roam-ci.yml got concurrency: groups with cancel-in-progress: true; 9 jobs across architecture-guardian / cga-attestation / roam-ci got explicit timeout-minutes (deliberate per-job pick, not a uniform default). publish.yml left without concurrency — never cancel publishes. Drive-bys captured: W1095 (publish.yml timeouts), W1096 (roam.yml template — sealed inline below).
• W1091 — roam runs verify --all LAW 4 fix on state=unsigned + state=key_missing. +6 LOC at src/roam/commands/cmd_runs.py:1050-1058 + 1 new test (+39 LOC) at tests/test_ledger_signing.py. Pre-fix: agent_contract.next_commands was empty on both unsigned + key_missing paths (W1007 surfaced this — LAW 4 violation per "Imperatives beat descriptions"). Post-fix: both branches now populate an imperative next_command (roam runs sign / roam runs verify --key-path <path>). Hash-stable for tampered + ok paths — the fix only adds bytes to envelopes that were previously omitting them. Drive-by W1097 (placeholder unify) captured.
• W1096 — roam.yml dormant template hardened. +1 line (timeout-minutes: 20) + 3-line teaching comment in the dormant templates/distribution/.../roam.yml template. Workflow-dispatch-only confirmed; template only fires on manual trigger. Mirrors the W1087 timeout-minutes discipline at the user-facing template surface.

Audits / NO-OPs — W1041-W1096 batch

• W1041 — clones_cross_layer.py __all__ already alphabetical (NO-OP). Verified the W1037 sweep convention matches W855/W856/W857/W858 sibling files. 101 focused tests pass. Drive-by W1090 captured (3 alphabetical-ordering conventions across 9 catalog files — narrow style-rule documentation candidate).
• **W1048 — actions/* sweep across .github/workflows/ — SWEPT-CLEAN.** All actions/* references already on current majors; this is a pure-Python repo so no setup-node versions to bump. Drive-bys captured: W1087 (shipped here), W1088, W1089.
• W1060-narrowed — cmd_complexity emit_runtime_notifications plumb — BAILED. Verified cmd_complexity has zero warnings_out accumulators — the proposed plumb would have been cargo-cult (no surface to plumb to). Prereqs captured: W1084 (cmd_health re-dispatch), W1085 (cmd_doctor SARIF surface), W1086 (warnings_out prereq across affected commands).
• W1007 — agent_contract:[] empty-list mistake — VALIDATED then SEALED by W1091. Confirmed cmd_runs.py:1050-1058 emitted empty next_commands on state=unsigned + state=key_missing. Tier-1 fix shipped via W1091 (above). Tier-2 design question captured as W1098-bis (auto-derive omit-when-empty across envelopes — DESIGN Q, not in-flight).
• W1008 — envelope-root list_counts sweep — BAILED. list_counts only exists as a dead local variable inside formatter.strip_list_payloads; already converged via W706+W1057. Drive-by W1093 captured (dead-code cleanup, deferred — formatter.py modified).
• W1004 — 7-cmd click-vocab audit — VALIDATED, DESIGN Q open. 6 commands (disambiguate / guard / safe_delete / symbol / test_scaffold / uses) diverge on @click.argument("name") vs the canonical MCP "symbol". Click has NO argument-alias support → migration approach captured as W1098 (USER DECISION pending). W1099 (path-cluster), W1100 (help-text), W1101 (memory plural) captured as siblings.
• W1005 — 3-tier vs 5-tier severity Pattern-3a — BAILED. Codebase already compliant with W547 + W564 discipline. CLAIM_SEVERITIES (5-tier evidence vocabulary) vs canonical 4-tier output vocabulary is layered by design — not a Pattern-3a divergence. Drive-by W1094 captured (docstring reconciliation).
• W1020 — fixture-scope audit — NO-OP. Already optimised: 8 test files use scope="module" override (~642s wall-clock savings); 6 findings test files cannot apply the override (DB mutations require per-test isolation).

Fixed — CI publish.yml (post-v13.1)

• W1047 — publish.yml SBOM-upload step fixed. The post-build gh release upload step now passes --repo explicitly AND creates the GitHub Release idempotently (gh release create ... || true) before attempting the upload, so the step no longer races the release-object creation. PyPI wheel content is unchanged — the v13.1 tag was force-moved to the fix commit (484e34fa) and the re-run went green end-to-end including the post-publish smoke (workflow run 25932785927, the first fully-green publish run in the history of the workflow). Both v13.0 and v13.1 GitHub Releases backfilled with their CycloneDX SBOMs as part of the same fix.

W1079-CONSOLIDATE — Pattern-1D + closest-match disclosure arc + helper-hoist Phase 1 + Pattern-2 propagation finale

> CONSOLIDATE checkpoint = W1079. ~17 waves closed since W1042-CONSOLIDATE. > Headline: the Pattern-1D unknown-value disclosure arc went from 1 > command (W1063 cmd_findings --detector) to 9 commands in a single > batch — cmd_findings / cmd_search --kind (W1068) / cmd_endpoints > --framework (W1069) / cmd_endpoints --method (W1075) / > cmd_test_scaffold --framework (W1070) / cmd_workflow + cmd_explain_command > (W1074) / cmd_oracle (W1079) / cmd_smells (W1066) — each emitting an > explicit structured envelope on the unknown-value path plus a > difflib.get_close_matches-derived "did you mean?" hint when the typo > distance is plausible. Pattern shape: closed enum → reject unknown > with state="unknown_<axis>" + partial_success=true + agent_contract.facts > listing the valid set + next_command carrying the closest match. > Mirrors the W918 / W994 / W995 / W1009 / W1011 / W1032 / W1042 loader > envelope shape — Pattern-1D ("silent success on degraded resolution" > from CLAUDE.md §"Six systemic anti-patterns") is now the 9th member > of the Pattern-2 propagation family. W1078 added a deliberate > click.Choice carve-out (cmd_complete --kind is click-validated, so > the Pattern-1D template does not apply; closed not-applicable). > Helper hoist (W1077) shipped src/roam/output/structured_unknowns.py::structured_unknown_filter > as Phase 1 (UNUSED on landing — 128 LOC + 15 tests; Phase 2 migration > W1080 in flight at consolidation time). Pattern-2 propagation > closures — W1010 final cmd_flag_dead._load_known_stale plumbing-only > close (plain-text loader, not YAML — does not flow through > load_yaml_with_warnings); W1043 WarningsOut type alias swept > 21 callsites across 8 files in one consistent application. Operational > findings: W1067 permit-expiry investigation closed NOT-A-BUG > (audit-completeness design per W377); W1071 documented permit-vs-lease > asymmetry in module docstrings + CLAUDE.md; W1072 — claude > subagent is structurally broken on Windows host via the W686 > worktree-MAX_PATH regression (the agent platform's default-worktree > behavior is the structural issue; the general-purpose subagent > works around it by not creating a worktree). W1076 documented that > CLAUDE.md is intentionally untracked (commit 89a338d9 removed it > from public repo — local-only by design). Test discipline: W1027 > extracted the _no_pyyaml monkeypatch to a tests/conftest.py > fixture (6 test files migrated, -50 LOC); W1059 converted 10 > hardcoded expires_at future-dates to relative offsets across 2 > files; W1065 triaged 3 more files with 0 conversions (all valid > B-variant expires_at-as-input fixtures). Research: W1049-RESEARCH > shipped a release-pipeline hardening memo with 3 P1 recommendations > (PEP740 attestations + workflow split + SBOM-wheel SHA binding) — > queued as W1054 / W1055 / W1056 user-decision-pending. Hash-stability > mandate held trivially across the batch — every Pattern-1D fix > added a new validation path (no pre-fix envelope bytes to compare); > W1077 helper shipped unused (no callsites yet); W1010 / W1043 / W1027 > / W1059 are docs/types/test-fixture only with no runtime behavior > delta. NO commits taken during the session per directive — entire > batch on the working tree for review.

Added — Pattern-1D + closest-match disclosure arc (W1066 / W1068 / W1069 / W1070 / W1074 / W1075 / W1078 / W1079 — W1079-CONSOLIDATE)

• W1068 — cmd_search --kind unknown-value disclosure + LAW 4 kinds anchor. Closed-set rejection of unknown --kind arguments now emits the canonical Pattern-1D envelope (state="unknown_kind" + partial_success=true + agent_contract.facts listing valid kinds + next_command carrying the difflib closest match). LAW 4 concrete-noun anchor set extended with kinds terminal in both src/roam/output/formatter.py:concrete_plural_terminals and tests/test_law4_lint.py:_CONCRETE_NOUN_ANCHORS to keep the lint compliant on the new fact strings.
• W1069 — cmd_endpoints --framework unknown-substring disclosure. Framework-name validation tightened from substring-match to exact-match against a closed framework registry; unknown values now emit the same Pattern-1D envelope with the supported framework list surfaced via agent_contract.facts. Closes a silent-fallback class where --framework djang was previously matching django and --framework f matched the first framework alphabetically.
• W1070 — cmd_test_scaffold --framework unknown-value disclosure. Sister fix to W1069 for the test-scaffold path; same Pattern-1D envelope.
• W1074 — cmd_workflow + cmd_explain_command UsageError difflib augmentation. Both commands previously raised click.UsageError on unknown workflow / command names with no suggestion. Fix: augment the UsageError message with the difflib closest match when distance ≤ cutoff. UsageError shape preserved (no envelope change) — the fix is purely the error message augmentation per LAW 2 imperative voice ("Did you mean roam <closest>?").
• W1075 — cmd_endpoints --method Pattern-1D disclosure. Method-name validation aligned with the W1069 framework path; closed HTTP-verb set, unknown values surface via the canonical envelope.
• W1078 — cmd_complete --kind audit closed not-applicable. cmd_complete --kind uses click.Choice(...) which already rejects unknown values pre-handler via the canonical Click error shape. No Pattern-1D template applies; documented as a deliberate carve-out so the next reader doesn't try to apply the W1068 pattern here.
• W1066 — cmd_findings + cmd_smells difflib closest-match augmentation. Both commands already rejected unknown values (post-W1063); this wave adds the difflib-derived next_command carrying the closest match. Pairs the W1063 disclosure shape with the W1074 difflib suggestion pattern.
• W1079 — cmd_oracle unknown-oracle name closest-match. Per-line shape: each unknown oracle name in the --oracles repeatable flag emits its own structured envelope row with state="unknown_oracle" + closest-match suggestion. Closes the final unknown-value site in the disclosure arc.

Added — Helper hoist Phase 1 (W1077 — W1079-CONSOLIDATE)

• W1077 — structured_unknown_filter helper shipped UNUSED (Phase 1). New module src/roam/output/structured_unknowns.py (128 LOC) + 15 tests in tests/test_structured_unknowns.py. Mirrors the W1018 shared YAML helper landing pattern (Phase 1 lands unused; Phase 2 migration proves the abstraction against real callsites). Phase 2 (W1080, in flight as of this consolidation) migrates the first 3 callsites — cmd_findings + cmd_search + cmd_endpoints --framework — and is expected to net -90 to -120 LOC once landed. The Pattern-1D template defined inline at each W1068 / W1069 / W1070 / W1075 / W1079 callsite is deliberately duplicated during Phase 1 to keep each fix reviewable in isolation; Phase 2 consolidates the duplication once 3 callsites have stabilised.

Pattern-2 propagation closures (W1010 / W1043 — W1079-CONSOLIDATE)

• W1010 final — cmd_flag_dead._load_known_stale plumbing-only Pattern-2 close. The W1015-batch deferred close (originally queued behind W1018) sealed here as a plain-text loader migration (NOT through load_yaml_with_warnings — the file is line-oriented, not YAML). warnings_out plumb + partial_success=true envelope flip on malformed lines + agent_contract.facts surface. Mirrors the W994 / W995 / W1009 shape but at the plain-text-loader boundary.
• W1043 — WarningsOut type alias applied across 21 callsites in 8 files. The W1042-CONSOLIDATE batch shipped WarningsOut: TypeAlias = list[str] at the canonical boundary; this wave swept the type alias through every callsite consuming warnings_out: list[str] | None (21 sites across 8 files). No runtime behavior change — readability + LSP hint quality only. Closes the W706-fan-out arc end-to-end that ran from W918 (cmd_alerts close) through the W994/W995/W1009/W1011/W1017/W1025/W1032/W1042 loader-site closures.

Test discipline (W1027 / W1059 / W1065 — W1079-CONSOLIDATE)

• W1027 — _no_pyyaml monkeypatch extracted to tests/conftest.py fixture. 6 test files migrated to consume the new shared fixture; -50 LOC net across the test corpus. Closes the duplicated-test-scaffolding class that the W934 catalog-finding-test parametrisation arc had been chipping away at.
• W1059 — 10 hardcoded expires_at future-dates converted to relative offsets across 2 files. Same shape as W1002 + W1012 (autouse freeze_time fixture interaction); preserves the W1002 / W1003 / W1012 discipline.
• W1065 — 3 more files triaged, 0 conversions. All 3 were valid B-variant test fixtures (the date is the INPUT being tested, not a free parameter); documented as a deliberate carve-out so the next sweep doesn't re-triage them.

Operational findings (W1067 / W1071 / W1072 / W1076 — W1079-CONSOLIDATE)

• W1067 — Permit-expiry investigation closed NOT-A-BUG. Reported as a permit-expiry race; investigation confirmed audit-completeness design per the W377 marker — permits intentionally remain in the audit trail post-expiry rather than being silently dropped. Documented inline in the permit module docstring.
• W1071 — Permit-vs-lease asymmetry documented. Sibling capture from the W1067 investigation; the permit + lease modules differ on expiry semantics in ways the design intends but the docstrings did not previously articulate. Fix: module-docstring updates on both surfaces + a CLAUDE.md sub-section codifying the asymmetry.
• W1072 — claude subagent structurally broken on Windows host (operational). The claude subagent creates a worktree by default; W686 / W903 captured the Windows MAX_PATH failure mode that makes this unusable on the roam-code repo path depth. Use the general-purpose subagent instead — it does not create a worktree. This is a structural issue with the agent platform's default-worktree behavior, not addressable from inside roam.
• W1076 — CLAUDE.md is intentionally untracked. Commit 89a338d9 removed CLAUDE.md from the public repo; the file is now local-only by design (private development guide; not shipped to PyPI/GitHub/landing-page). Documented inline at the top of CLAUDE.md so future readers don't try to re-add it to git.

Research memos (W1049-RESEARCH — W1079-CONSOLIDATE)

• W1049-RESEARCH — Release-pipeline hardening memo, 3 P1 recommendations. Post-W1047 audit of the publish.yml workflow surfaced three P1 hardening recommendations: (1) PEP 740 attestations — sign the wheel + sdist with provenance attestation post-OIDC-mint and attach to the GitHub Release alongside the SBOM; (2) workflow split — separate build / publish / smoke into independent workflows so a smoke-step failure doesn't block the publish from finishing; (3) SBOM-wheel SHA binding — the CycloneDX SBOM should carry the wheel's content SHA so the SBOM cannot be silently swapped post-publish. Queued as W1054 / W1055 / W1056 — user-decision-pending per the v13.1 release-prep state CLOSED gate.

v13.1 (released 2026-05-15) -- Pattern-2 propagation + shared YAML helper + 3 flagship silent-fallback seals

> THREE flagship Pattern 2 silent-fallback bugs SEALED this batch (W826 cmd_taint + W834 cmd_health + W836 cmd_doctor) + W817 helper-level auto-inject closed Pattern 2 partial_success gap across 7 detectors in one shot (dead / clones / complexity / orphan-imports / bus-factor / auth-gaps / hotspots) + W810 cmd_complexity Pattern 1B fix (SystemExit(1) → return on empty corpus) + W805 empty-corpus sweep covered 25+ detectors (cmd_endpoints/n1/missing-index/over-fetch/smells/duplicates/invariants/vulns/audit-trail-conformance/audit-trail-verify/pr-risk/critique already clean; 7 auto-fixed by W817; 3 flagship dedicated fixes) + W749 dispatch-edge MIN(id) fix in registry_dispatch (231 + 34 + 22 edge attribution corrections) + W774 sister fix to laravel_post.py (worktree-pending) + W718 cleaned 70+ UPPER-case severity sites + W634 confidence_level_rank fail-loud + 15 callers + W444 mcp_tool_names duplicate fail-loud + W445 _REGISTERED_TOOLS append guard + W707 _serialize_suppressions dead-code seal + drift-guard expansion (W703 _CommentSyntax + W741 find_project_root symlink-safety + W484 templates/ci/ reachability + W711/W712 mcp --card/--list-tools coverage + W713 _SARIF_CONSUMERS AST-literal + W757 backfilled missing W702) + W397 build_readme_counts AGENTS.md + W734 CONTRIBUTING.md count refresh + five research memos shipped in dev/ (MCP-EVOLUTION / MCP-SERVER-CARD / MCP-TASKS-EVAL / MCP-ELICITATION-CANDIDATES / DETECTOR-FP-METHODOLOGY) (~50-completion batch behind W836-CONSOLIDATE, 2026-05-15). > The headline is THREE flagship Pattern 2 silent-fallback bugs sealed in main this batch — all three share the "claim success on unanalyzed corpus" shape that the W805 empty-corpus sweep was designed to surface. Mirroring the proven cmd_vulns Fix E template (state="no_scan" + partial_success=True + actionable verdict) and the cmd_missing_index state="no_migrations" discipline, each fix detects the empty-graph precondition BEFORE running the rule/check pipeline and emits an explicit Pattern-2 envelope instead of a default-success illusion. W826 cmd_taint — previously emitted "No taint findings across 22 rule(s)" + partial_success=false (W817 auto-injected, making the false claim deterministic) on a fully-empty corpus; the verdict read as a clean security pass on an unanalyzed repo. Fix: 52-line empty-corpus guard right after open_db(...); on COUNT() FROM symbols == 0 emits state="empty_corpus", partial_success=true, rules (count loaded but not run), verdict "no symbols to analyze (corpus empty; N rules loaded but not run — run \roam index --force\ to populate the graph)". Test tests/test_w825_taint_empty_corpus.py xfail-strict flipped green plain; 35/35 taint regression tests pass. W834 cmd_health — the FLAGSHIP CI-gate command previously emitted verdict: "Healthy codebase (100/100) — 0 critical issues" with health_score: 100 on an empty corpus, because every health factor defaulted to 1.0 on zero signal and the geometric mean returned exactly 100 → score ≥ 80 threshold matched "Healthy codebase". A 100/100 Healthy verdict on an unanalyzed repo was a HIGH-severity false claim that would silently pass CI gates. Fix: 65-line empty-corpus carve-out before build_symbol_graph(...); emits state="empty_corpus", partial_success=true, health_score=None (not 0, not 100), verdict "no symbols to analyze" + next_command="roam index --force". --gate flag raises GateFailureError (exit 5) on empty corpus — mirrors W829 audit-trail-verify discipline (fail-closed on missing analysis). Test 1/1 pass; 93/93 health regression tests pass; LAW 4 lint 8/8 pass. W836 cmd_doctor — previously checked only environment markers (Python/tree-sitter/git/networkx/manifest) and never asked "did the indexer extract anything?". On a clean env + empty corpus, would emit "all N checks passed" even though zero symbols had been indexed. Fix: new _check_corpus_content() function queries SELECT COUNT() FROM symbols; states no_index / empty (advisory fail with actionable verdict) / populated / error. Wired into the check pipeline after _check_required_tables and into _ADVISORY_CHECK_NAMES so empty corpus warns but does not block CI by default. Total check count bumped 23 → 24; 71/71 doctor regression tests pass; 5/5 W835 tests pass. W817 helper-level auto-inject closes Pattern 2 across 7 detectors in one shot: added a 9-line auto-inject at src/roam/output/formatter.py:json_envelope() defaulting summary.partial_success to False when missing. Closed the gap for 7 detectors without per-command edits; companion W819 manual xfail-strict flip on the 7 corresponding empty-corpus smoke tests; W810 manual cmd_complexity Pattern 1B fix (SystemExit(1) → return). W749 dispatch edge-attribution chain extended: discovered dispatch edges were 100% mis-attributed via a DIFFERENT mechanism than W742 — MIN(id) synthetic-source in registry_dispatch.py. Replaced with per-file [(line_start, line_end, symbol_id)] map + _symbol_for_assignment lexical-extent lookup. _COMMANDS now sources 231 dispatch edges (was attributed to _DEPRECATED_COMMANDS); _MATH_DETECTORS 34 edges (was attributed to log); PYTHON_IDIOM_DETECTORS 22 edges (was attributed to _MUTABLE_DEFAULT_RE). W805 empty-corpus sweep methodology validated: 25+ detector commands smoke-tested; 12 already Pattern-2 clean; 7 auto-fixed by W817; 3 flagship dedicated fixes; every smoke test ships with a forbidden-fragment blacklist ("safe" / "healthy" / "no concerns" / "all clear" / "100/100") as a regression guard. W772 worktree-staleness operational finding: ~8 dispatches bailed because agent worktrees branch from commit 850552af (pre-session main); user's git config --global core.longpaths true fix resolved the parallel W686 path-length issue. Hash-stability mandate held across every fix — tests/test_evidence_schema_migration.py 31/31 byte-identical.

Added — Detector inventory memo (W850)

• W850: dev/ROAM-DETECTOR-INVENTORY-2026-05-15.md <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->

Fixed — THREE flagship Pattern 2 silent-fallback bugs SEALED (W826 taint + W834 health + W836 doctor)

• W826 — HIGH-SEV: cmd_taint silent-SAFE on empty corpus sealed. Added 52-line empty-corpus guard at src/roam/commands/cmd_taint.py after open_db(...). On COUNT(*) FROM symbols == 0, emits Pattern 2 envelope (state="empty_corpus", partial_success=True, rules count loaded-but-not-run, actionable verdict). Mirror of cmd_vulns Fix E. tests/test_w825_taint_empty_corpus.py xfail-strict flipped green (2/2); 35/35 taint regression tests pass. Security-critical — pre-fix verdict read as a clean security pass on unanalyzed code.
• W834 — CRITICAL: cmd_health silent-Healthy 100/100 on empty corpus sealed. 65-line empty-corpus carve-out at src/roam/commands/cmd_health.py before build_symbol_graph(...). Geometric mean of health factors (each 1.0 on zero signal) was deterministically returning 100, triggering "Healthy codebase" threshold. Fix emits state="empty_corpus", partial_success=True, health_score=None, verdict "no symbols to analyze", next_command="roam index --force". --gate flag raises GateFailureError (exit 5) on empty (W829 fail-closed discipline). 1/1 pass + 93/93 regression + 8/8 LAW 4 lint. FLAGSHIP CI-gate bug — 100/100 Healthy on unanalyzed repo would have silently passed every roam health --gate CI check.
• W836 — HIGH: cmd_doctor silent "all checks passed" on empty corpus sealed. New _check_corpus_content() in src/roam/commands/cmd_doctor.py (~70 lines). States: no_index (install OK, advisory pass) / empty (advisory fail) / populated / error. Wired into pipeline after _check_required_tables + _ADVISORY_CHECK_NAMES. Verdict computation extended with explicit "corpus empty (0 symbols)" branches. Total check count 23 → 24; 71/71 doctor regression tests pass.
• W817 — helper-level Pattern 2 closure. 9-line auto-inject at src/roam/output/formatter.py:json_envelope() defaulting summary.partial_success to False when missing. Closes the gap for 7 detectors (cmd_dead / cmd_complexity / cmd_clones / cmd_orphan_imports / cmd_bus_factor / cmd_auth_gaps / cmd_hotspots) in one edit. Companion W819 manual xfail-strict flip on the 7 empty-corpus smoke tests.
• W810 — cmd_complexity Pattern 1B fix. Empty-corpus branch was emitting structured envelope + raise SystemExit(1) — Pattern 1 variant B (wrapper-bridge converts to generic COMMAND_FAILED). Fix: changed to clean return, added state="no_complexity_data" + next_command.

Fixed — Edge-attribution chain extended (W749 + W774)

• W749 — Dispatch edges 100% mis-attributed via MIN(id) in registry_dispatch.py — SEALED. Different mechanism than W742. Replaced with per-file [(line_start, line_end, symbol_id)] map + _symbol_for_assignment lexical-extent lookup. _COMMANDS now sources 231 dispatch edges (was attributed to _DEPRECATED_COMMANDS); _MATH_DETECTORS 34 edges (was attributed to log = logging.getLogger); PYTHON_IDIOM_DETECTORS 22 edges (was attributed to _MUTABLE_DEFAULT_RE). 106/106 focused tests pass; hash-stability preserved.
• W774 — laravel_post.py MIN(id) anti-pattern fix. Same root cause as W749. Applied analogous fix using a new shared helper src/roam/index/_containing_symbol.py (build_file_symbol_ranges() + containing_symbol_for_line()). 129/129 focused tests pass. (Note: this work landed in a worktree branch on the pre-W624 baseline — pending careful cherry-pick to main via W782 because direct copy would regress mcp_server.py importlib.resources migration.)

Fixed — W805 empty-corpus sweep: methodology + 25+ detectors smoke-tested

• Already Pattern-2 clean (regression guards shipped only): cmd_endpoints (W801), cmd_n1 (W803), cmd_missing_index (W807 — state="no_migrations"), cmd_over_fetch (W809), cmd_smells (W820), cmd_duplicates (W821), cmd_invariants (W824), cmd_vulns (W823 — Fix E pre-existing), cmd_audit_trail_conformance (W827 — Fix E + Article 12 not_run), cmd_audit_trail_verify (W829 — 3-state matrix), cmd_pr_risk (W828), cmd_critique (W831 — structured EMPTY_INPUT: UsageError).
• Auto-fixed by W817 helper (xfail-strict flipped green): cmd_dead (W802/W804), cmd_complexity (W806 + W810 Pattern 1B), cmd_clones (W808/W813), cmd_orphan_imports (W812/W814), cmd_bus_factor (W811/W817), cmd_auth_gaps (W815/W818), cmd_hotspots (W816).
• Three flagship surfaces dedicated-fixed: cmd_taint (W826), cmd_health (W834), cmd_doctor (W836) — see preceding section.
• World-model R28 classifiers (W680): empty-corpus smoke for side-effects/idempotency/causal-graph/tx-boundaries; all 4 already compliant.
• taint_engine positive smoke (W681): 3 tests proving SSTI co-call + SQLi forward BFS + shipped-YAML rule-pack matching.
• Forbidden-fragment blacklist discipline (W823 / W825 templates): every new smoke test includes a verdict blacklist preventing future Pattern 2 regressions.

Fixed — Defensive fail-loud expansion (W444 / W445 / W634 / W707)

• W444 — mcp_tool_names() fail-loud on duplicates. src/roam/surface_counts.py:mcp_tool_names() historically returned sorted(set(names)) silently collapsing duplicate @_tool(name=...) decorations. Now raises ValueError with offending names. 14 callers audited + classified; 3-test smoke harness pins the contract.
• W445 — Defensive duplicate-check in _REGISTERED_TOOLS.append. 9-line guard at src/roam/mcp_server.py:872 raising RuntimeError("Duplicate MCP tool registration: ..."). Runtime + AST defense-in-depth against W432-class duplicate-registration bugs.
• W634 — confidence_level_rank() flips to fail-loud by default. Added fallback: int | None = None kwarg. Raises ValueError on unknown without fallback. 15 callers migrated: 13 explicit fallback=-1 (data-path); 2 fail-loud (CLI-validated input). 57 + 128 focused tests pass.
• W707 — REAL BUG: _serialize_suppressions dead-code on first flag. Dead flag set but never read; removed; regression test pins call-site count at zero.

Changed — UPPER-case severity vocabulary canonicalisation (W632 / W718)

• W718 — cleaned 70+ UPPER-case severity sites across cmd_health (50), cmd_pr_risk (8), cmd_path_coverage (12). Every envelope severity field, findings-registry row, SARIF input, text formatter call now lowercase per W547 contract. 31/31 hash-stability + 324 focused + 77 critique tests pass.
• W632 — cmd_path_coverage UPPER-case risk 4-tier canonicalised. Single comment-only stale reference; underlying code was already W718-canonical.

Added — Coverage tests + drift guards (W397 / W444 / W484 / W680 / W681 / W703 / W711 / W712 / W713 / W734 / W741 / W757 / W801–W835)

• W397 — build_readme_counts.py AGENTS.md integration + count refresh. New _agents_md_blocks() builder; refreshed Codex-headline (233 → 238 commands, 57 core / 149 full → 224 MCP tools (57 core)) + Codex-authoritative (canonical 226 → 231) to current canonical counts.
• W484 — templates/ci/ wheel-bundling reachability test. 9 tests pinning importlib.resources.files("roam.templates.ci") reachability + __init__.py marker per W664 discipline.
• W703 — _CommentSyntax language-coverage drift-guard. 4 tests pinning every canonical language is in _COMMENT_SYNTAX_BY_LANG (30 entries) OR in _COMMENT_DENSITY_NO_SUPPORT skip-set. Added 7 languages (tsx/jsonc/vue/svelte/sfxml/aura/visualforce).
• W711 — mcp --card error-branch coverage. 4 tests covering FileNotFoundError / corrupted-JSON / OSError / no-traceback-leak. Drive-by found W788 (handler not Pattern-1 canonical) + W789 (hash-pin drift).
• W712 — mcp --list-tools success-path coverage. 5 tests pinning exit-0 + canonical core tool names + JSON envelope shape + presets advertisement.
• W713 — _SARIF_CONSUMERS AST-literal contract test. 2 tests asserting constant is ast.Tuple of ast.Constant strings AND parsed tuple equals runtime value.
• W741 — find_project_root .git symlink-safety regression test. 4 tests pinning that worktree-pointer-FILE wins over a parent real .git directory.
• W734 — CONTRIBUTING.md count refresh. 3 stale references fixed: rev: v11.1.2 → v13.0, MCP server with 101 → 224 tools (57 in core preset), Test suite (186 → 408 test files).
• W757 — Backfilled missing W702 _DEPRECATED_COMMANDS schema test. W702 was a false-completion in TaskList. Wrote tests/test_cli_deprecated_commands_schema.py (5 AST-literal schema tests, all pass).
• W680 — Empty-corpus smoke for world_model R28 classifiers (4 tests, all clean).
• W681 — Positive smoke for security/taint_engine (3 tests proving real SSTI + SQLi taint flows).
• W801 – W835 — Per-detector empty-corpus smoke tests (~20 new files, one per detector; see W805 sweep section).

Research memos shipped (5 in dev/)

• dev/MCP-EVOLUTION-2026-05-15.md (W758) — MCP protocol evolution since 2025-11-25 spec freeze. Headline: Claude Code bug #25081 silently drops ALL tools from any server that declares toolAnnotations or outputSchema — hard block on tool-annotation adoption (W764 / W767 BLOCKED). 5 follow-up waves suggested.
• dev/MCP-SERVER-CARD-2026-05-15.md (W765) — MCP Server Card readiness vs SEP-2127 draft. 7 gaps; 4 ready (W792-W794); 1 BLOCKED on SEP-2127 merge (W795). 10 sources cited.
• dev/MCP-TASKS-EVAL-2026-05-15.md (W766) — MCP Tasks primitive evaluation. CRITICAL: 8 sync tools falsely declared in _TASK_OPTIONAL_TOOLS (W785 captured). 10 candidate tools; 6 prerequisites; 3-tool migration order.
• dev/MCP-ELICITATION-CANDIDATES-2026-05-15.md (W768) — Top-3 elicitation candidates: roam_mutate, roam_cga_emit, roam_annotate_symbol. 4 prerequisites; 4 runner-up candidates. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• dev/DETECTOR-FP-METHODOLOGY-2026-05-15.md (W480) — FP-rate methodology: hybrid (OWASP Benchmark / public corpora / internal dogfood). CRITICAL: 93% of BigCloneBench's Weak Type-3/4 pairs are mislabeled per arXiv 2505.04311 (W797 captured). 13 sources cited. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->

Operational findings (pending fix / decision)

• W797 — CRITICAL caveat: BigCloneBench Type-3/4 93% mislabeled. Any clones-detector FP-rate claim citing BigCloneBench must exclude Type-3/4 OR cite the caveat.
• W785 — CRITICAL: 8 sync tools falsely in _TASK_OPTIONAL_TOOLS. Convert to async def via _run_roam_async OR drop from the set. Add inspect.iscoroutinefunction() drift-guard.
• W789 — REAL BUG: mcp-server-card.json hash-pin drift. _EXPECTED_CARD_SHA256 not bumped after edit. W563 auto-rotate didn't fire — investigate why.
• W791 — REAL BUG: stale-index envelope drift in test_mcp_server.py. 2 tests assert exact-equality but runtime now decorates with _meta.stale_index + verdict suffix.
• W772 — Worktree-staleness pattern (recurring). ~8 dispatches bailed because agent worktrees branch from 850552af (pre-session main). User's git config --global core.longpaths true fix resolved the parallel W686 path-length issue.
• Three pending-merge worktree branches (W774 / W706 / W788) — completed work in stale worktrees; need cherry-pick.
• W798 — Click 8.3 dropped CliRunner(mix_stderr=False). Mass-replace required.

Added — Smell catalog detector roster 20 → 24 (W852 / W853 / W855 / W856 / W857 — W865-CONSOLIDATE)

• W853 — speculative-generality (YAGNI). New detector in src/roam/catalog/smells.py: flags symbols whose only callers are test files (test_.py / _test.py) — surfaces YAGNI scaffolding that exists solely to be unit-tested. Confidence tier structural; wired into ALL_DETECTORS.
• W857 — parallel-hierarchy (Fowler). New detector module src/roam/catalog/parallel_hierarchy.py (detect_parallel_hierarchy(conn) -> list[dict]) re-imported into smells.py. Mirrored-subclass-hierarchy smell: when adding a subclass on one side forces an analogous subclass on the other. 16/16 tests pass (tests/test_w857_parallel_hierarchy.py).
• W855 — Rename-invariant clones (DECKARD-style). New module src/roam/catalog/clones_rename_invariant.py shipping a characteristic-vector clone detector that survives identifier renames. 6/6 tests pass (tests/test_w855_rename_invariant_clones.py). 6,070 Type-2 pairs surfaced on roam-code itself. Library-layer only this batch — no CLI surface yet; the persistence + CLI wrapper is deliberately separated to keep the algorithm landing reviewable.
• W852 — type-switch (OCP / Fowler). New module src/roam/catalog/type_switch.py re-imported into smells.py. Detects chained isinstance / type(...) == / match-case dispatch against ≥3 concrete classes on a single discriminator — recommends Strategy / Visitor / singledispatch. All tests pass (tests/test_w852_type_switch.py).
• W856 — cross-layer-clone (the #1 real-world DRY debt class). New module src/roam/catalog/clones_cross_layer.py re-imported into smells.py. Implementation: Jaccard similarity over callee-NAME multisets across detected layers (controllers / services / repositories). Targets the strategic DRY debt — duplicated domain logic routed through different layers — that literal-clone detectors miss. 23/23 tests pass (tests/test_w856_cross_layer_clones.py). 0 findings on roam-code itself, which is correct: roam is a CLI library, not a layered web app.
• Smells.py wiring. Docstring count + run_all_detectors() docstring + ALL_DETECTORS list bumped 20 → 24 in lockstep. Module-level imports added for parallel_hierarchy, type_switch, clones_cross_layer. All five new test files green via pytest tests/test_w85{2,3,5,6,7}_*.py -x -q.

Added — Strategy memos shipped earlier in the W836→W865 arc (W848 / W849 / W850 / W859)

• W848 — dev/PRINCIPLE-ENFORCEMENT-ROADMAP-2026-05-15.md. Fowler 22-smell coverage map + top-3 recommendations for the next detector wave. Drove the W852-W857 selection.
• W849 — dev/DRY-BEYOND-LITERAL-CLONES-2026-05-15.md. Research memo identifying cross-layer duplication as the #1 real-world DRY debt class. Directly drove the W856 design (Jaccard over callee-name multisets across layers, NOT raw clone detection).
• W850 — dev/ROAM-DETECTOR-INVENTORY-2026-05-15.md. 94 distinct detectors catalogued across the codebase. The first single source of truth for the full detector roster. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• W859 — Correction banner on W848. W848's draft claimed empty-catch was a stub; W370 had already shipped a real detector. Banner added to the memo flagging the inventory drift — sealed silently before any downstream reader could be misled.

Operational findings (W865 batch — pending fix / decision)

• W862 — smells.py docstring count drift latent risk. Docstring "20 deterministic detectors" was silently allowed to drift to "24" while ALL_DETECTORS grew underneath. Both are now in sync at 24 but the discipline gap remains. Follow-up: add tests/test_smells_detector_count_matches_docstring.py that AST-parses the module docstring + asserts len(ALL_DETECTORS) == <docstring count>.
• W863 — ALL_DETECTORS entries ordered ad-hoc. Smells are appended in arrival order, not alphabetical-by-smell-id. If a future SARIF emitter ever depends on stable run-to-run ordering (e.g. for golden-fixture hashing), this becomes a real bug. Follow-up: standardize alphabetical-by-smell-id + AST drift-guard.
• W864 — _loc() helper duplicated 3 ways. Identical _loc() function definitions in src/roam/catalog/smells.py + src/roam/catalog/parallel_hierarchy.py + src/roam/catalog/clones_cross_layer.py — itself a W95-style clone in the detector code that's catching clones. Follow-up: hoist into src/roam/catalog/_shared.py.
• W861 — Worktree-isolation files surface on main (W783 follow-up). Files created by isolation: worktree agents appear in main as untracked rather than staying in the worktree. Not a bug per se, but worth documenting: future planning should treat "worktree isolation" as effectively "work in main + auto-stage" for additive changes. Pure code-isolation guarantees only hold for paths the agent does not modify.

Added — Catalog helper-hoist arc + registry-parity backstops (W886-CONSOLIDATE)

• W864 — src/roam/catalog/_shared.py created (~50 lines). Collapsed 4 _loc() definitions → 1 canonical + 2 _find_workspace_root() definitions → 1 canonical, across smells.py + clones_cross_layer.py + type_switch.py + detectors.py. 332 focused tests green. Closes the W95-style clone inside the detector code that catches clones.
• W873 — is_test_path() extension to _shared.py (~70 lines + 4 pattern tuples). Covers Python / Go / JS-TS / Java-Kotlin / Ruby / Apex test-naming conventions. Folded 2 catalog-layer duplicates: detectors._is_test_path (37 call-sites; _INCLUDE_TESTS_OVERRIDE semantics preserved) + type_switch._file_is_test (1 call-site). 6 non-catalog sites left alone (already delegated correctly OR canonical at their own layer OR deliberate import-cycle break). 17/17 new + 216/216 sibling tests pass.
• W862 — tests/test_smells_detector_count_drift.py (173 lines, 3 tests). AST-parses both the module docstring and run_all_detectors() docstring; asserts both stay in lockstep with len(ALL_DETECTORS). Catches the exact count-drift class W856 surfaced. Inline drift fixes: smells.py:2893 "remaining 19 detectors" → "remaining detectors"; cmd_smells.py:72 "The 15 detectors" → "The detectors".
• W867 — tests/test_smells_confidence_mapping_parity.py (208 lines, 3 tests). AST parity lint between ALL_DETECTORS ids and _SMELL_KIND_TO_CONFIDENCE keys. Reference set computed as ALL_DETECTORS ∪ AST-derived _finding("<id>",...) first-args so the W647 rollup pattern (one detector emits two smell_ids like temporal-coupling-cluster) doesn't false-trip.
• W869 — dev/REGISTRY-PARITY-PATTERN-2026-05-15.md (~600 lines, research memo). Synthesises the registry-parity bug-class across 10+ session-observed drift instances (W852 / W856 / W862 / W867 / W432 / W702 / W785 / W332 / W397 / W37.1-W113); 8 industry references. Recommendation: hybrid Archetype B+E (decorator-driven @detector(smell_id, confidence_tier) + construction-time validation + parity-test backstop). P0 = smell-detector registry (24 detectors, W871); P1 = MCP tool registry (224 tools, next sprint); P2 = mode-allowlists / _DEPRECATED_COMMANDS / subject_kind (rare-touch). <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• W874 — "Mirror smells.py" docstring anti-pattern sweep. Audited ~75 mentions; only 4-5 are real code clones (captured as W877-W880 drive-bys). The rest are legitimate comparative-narrative documenting parallel-but-distinct behavior.
• W876 — Stale-pending triage cleanup. Audited 20 candidates from the long-running BACKLOG queue; flipped 13 actual stale-pending rows pending → shipped: W125 (Wave30.1 doc-hygiene — already shipped W250; wave-number-vs-task-ID collision documented), W221 (flagged user-blocked), W224 (superseded by W240 / W242 / W261 / W266 / W267 / W268 producer-sealing waves), W298-polish, W319, W335, W342, W345, W346, W348, W349 (W377-W382 closed by W436 batch), W352, W353. W107 left pending — mode taxonomy still awaits user signoff.

Operational findings (W886 batch — pending fix / decision)

• **W870 — Per-detector *_DETECTOR_VERSION sparse-stamp lint candidate** (W867 finding). Some detectors stamp <NAME>_DETECTOR_VERSION per the Adding-a-command checklist; others don't. Worth a parity lint asserting every ALL_DETECTORS entry declares one.
• W871 — P0 @detector decorator implementation (W869 recommendation). Smell-detector registry is the P0 surface for the hybrid Archetype B+E rewrite. ~24 detectors to migrate; backstop tests (W862 + W867) already in place.
• W872 — Layer-classification heuristic audit across clone detectors (W864 finding). Controller/service/repository inference heuristics inside the clone detectors may have similar drift to the now-folded _find_workspace_root — audit candidate.
• W875 — Consolidate _finding / _make_finding constructors across smells.py + detectors.py. Two near-identical constructors for the same payload shape; candidate for _shared.py extension.
• W877-W880 (W874 findings) — Real clones outside the catalog layer. Hoist _enclosing_symbol from type_switch.py (W877); fold _bare_command_name triple-mirror across cmd_next.py + constitution/loader.py + modes/policy.py (W878); hoist _camel_split from retrieve/seeds.py (W879); fold _parse_iso duplication in change_evidence.py / approval.py (W880). All blocked behind W885.
• W881-W884 — Delegate _is_test_path in 4 non-catalog sites to canonical changed_files.is_test_file: cmd_over_fetch.py (W881), metrics_history.py (W882), rules/builtin.py (W883), rules/dataflow.py (W884). These are the sites W873 deliberately left alone pending W885.
• W885 — Architectural decision outstanding: extend changed_files.is_test_file coverage to fold W881-W884 in OR invent a roam._common namespace? W873 punted; W869 registry-parity logic favors extending the existing canonical site for single-source-of-truth.

Added — Cross-layer hoist execution + decorator-driven registry POC (W908-CONSOLIDATE)

• W877 — _enclosing_symbol hoisted to src/roam/catalog/_shared.py. Two near-identical sibling definitions (smells.py defensive + type_switch.py permissive) collapsed to a single canonical site; the defensive variant won (preserves try/except OperationalError contract — see W888 audit). type_switch.py local def removed + Mirror admission docstring stripped. 253 focused tests pass. Closes the W874-finding clone that lived inside the detector code that catches clones.
• W879 — _camel_split hoisted from retrieve/seeds.py → search/index_embeddings.py canonical (12 lines removed). Companion W901: added __all__ = ["_camel_split"] to index_embeddings.py to declare the underscored name as an intentionally exported cross-package boundary.
• W880 — _parse_iso hoisted from evidence/change_evidence.py → evidence/approval.py canonical (12 lines removed). The "duplicated here to avoid import cycle" docstring was VERIFIED FACTUALLY WRONG — no cycle exists. W902 captures this as a false-hedge pattern audit finding.
• W881-W884 — Bundle: 4 cross-layer _is_test_path delegations to canonical roam.commands.changed_files.is_test_file. 9 call-sites across cmd_over_fetch.py (W881), metrics_history.py (W882), rules/builtin.py (W883), rules/dataflow.py (W884) now route through the canonical. The canonical already covered every pattern the 4 sites needed — no extension required. 477 focused tests pass. This empirically RESOLVES W885 (the architectural decision punt): catalog/_shared stays narrow, cross-layer routes through changed_files, no transverse roam._common namespace needed.

Added — Decorator-driven smell-detector registry POC (W871)

• W871 — P0 decorator POC for the W869 registry-parity recommendation. New module src/roam/catalog/registry.py (176 lines) exposes @detector(smell_id, confidence_tier=...) + register_rollup_kind(...) + construction-time validation against the W867 vocabulary. 2 detectors migrated: speculative-generality (W853 detector) + temporal-coupling parent (W602) with temporal-coupling-cluster rollup (W647). 212 focused tests pass. Validates the hybrid Archetype B+E approach on a narrow surface before the bulk migration; remaining 22 detectors stay hand-rolled until the design follow-ups (W895 rollup_id auto-infer / W896 stable ordering / W897 parent_id finalization) close.
• W894 — Confidence-tier mismatch surfaced + sealed by W871. The decorator POC initially read temporal-coupling as structural because the rollup pattern in _SMELL_KIND_TO_CONFIDENCE indexed both parent + cluster at the same tier. Audit confirmed the HAND-ROLLED side was correct (W602 + W647 intentional split: parent = heuristic for git-cochange frequency, cluster rollup = structural for graph aggregation). Decorator side aligned. W867 lint extended with a new value-parity test so the same drift can't recur silently the next time a detector is migrated.

Fixed — Cross-language test-path false-positive (W889)

• W889 — Catalog is_test_path missed camelCase Java/Kotlin/C#/Swift/PHP/Scala/Apex Test/Tests basenames. Added 3 new case-sensitive regex patterns mirroring DEFAULT_TEST_PATTERNS so FooTest.java, BarTests.cs, BazTest.kt etc. are now recognised by the catalog-layer canonical. The W886 "we miss these" pinning test (originally written xfail-strict to document the gap) inverted into an 11-case positive + 8-case negative parametrize. 276 tests pass. Closes the W886 drive-by-1 cross-language gap.
• W891 — _TEST_FILE_SUFFIXES extended with _test.exs (Elixir) + _test.dart (Dart) for canonical parity with the W873 baseline. 334 tests pass.
• W893 — VERIFIED FALSE POSITIVE. The W889 follow-up flagged Apex _Test.cls as a coverage gap, but the canonical regex ^.Test\.cls$ already matches (greedy .* consumes the underscore). Pinned with a new 4-layer parity test asserting all suffix variants match across the canonical + catalog + changed_files + DEFAULT_TEST_PATTERNS layers.

Hardened — Stale-pending triage + drive-by audits (W902)

• W902 — Audited 6 "duplicated here to avoid X" docstrings across the codebase. Outcome: 1 real-now-resolved (W880); 3 FALSE HEDGES captured as new pendings (loader.py / django_post.py / cmd_oracle.py — claims of import-cycle avoidance that are factually wrong); 2 forward-looking kept (mcp_server.py / oscal.py — defensive lazy-import comments that document genuine optional-dep handling). Meta-observation captured as W907: the false-hedge cargo-cult anti-pattern was replicated 3+ times, suggesting a CLAUDE.md note is worth landing the next time the LLM-discipline section gets touched.
• W876 follow-on stale-pending flips (continuing the W886 cleanup): 11 additional rows flipped pending → completed via the W876 methodology — W107, W125 (false-positive), W221, W224, W298-polish, W319, W335, W342, W345, W346, W348, W349, W352, W353. W107 specifically was confirmed previously-shipped — the user-blocked gating note was stale.

Operational findings (W908 batch — pending fix / decision)

• W887 — python_idioms._enclosing_symbol name collision (W877 drive-by). A third site exists under src/roam/python_idioms/ that wasn't part of the W877 hoist — name collision rather than clone; audit candidate for naming convention disambiguation.
• W888 — smells._enclosing_symbol defensive-migration audit (W877 drive-by). Confirmed the defensive variant (the one with try/except OperationalError) is correct as canonical; sister sites silently degrading to syms[0] is the bug pattern this guards against.
• W895 / W896 / W897 — W871 decorator follow-ups. @detector POC works on 2 detectors but the bulk migration is blocked behind three design decisions: (W895) rollup_id auto-infer from parent + suffix vs explicit kwarg; (W896) stable iteration ordering — alphabetical-by-smell-id vs declaration-order frozenset; (W897) parent_id finalisation semantics for the rollup pattern. None blocking; each is a 30-min design call.
• W898 — Long-term catalog/_shared.is_test_path delegate to canonical. Sister candidate to W885 — _shared.is_test_path could itself delegate to changed_files.is_test_file instead of carrying its own regex tuples. Defers cleanly behind the W871 bulk migration so the registry POC has a stable surface.
• W899 — Tighten the Apex Test.cls regex. W893 confirmed the greedy .* works but is non-obvious; an explicit (?:_)?Test alternation would document intent. Cosmetic but documentation-grade.
• W900 — Per-language adapter table (suffix-tuple + camelCase-pattern-tuple unification across the 4 layers W893 surfaced). Today each layer carries its own pattern tuples; a single per-language adapter table would centralise the discipline. Deferred behind W898 so the canonical site moves first.
• W903 — CRITICAL operational: claude subagent type creating worktrees by default → W686 path-length blocking. ~3 dispatches this batch failed because the agent worktrees branched off old SHAs with paths exceeding the Windows MAX_PATH limit despite the prior session's git config --global core.longpaths true fix. Captured for a tooling-side investigation; not addressable from inside roam.
• **W904 — django_post._DJANGO_* constants duplicated from python_lang** (W902 finding). Trivial hoist — same shape as the W879/W880 pattern, no architectural decision required.
• W905 — cmd_oracle.py:83 lazy-import false-hedge claim (W902 finding). Comment claims import-cycle avoidance; no cycle exists. Cosmetic cleanup.
• W906 — Overly-defensive lazy-import comments in mcp_server.py + oscal.py (W902 forward-looking). Kept as-is for the W908 batch — both surface real optional-dep handling but the wording reads as cargo-cult; worth a polish pass in a future docs-only wave.
• W907 — CLAUDE.md note on the false-cycle hedge cargo-cult anti-pattern (W902 meta-observation). The pattern has replicated 3+ times across unrelated modules; a one-paragraph note in CLAUDE.md's LLM-discipline section would deter the next replication.

Fixed — Registry-parity remediation HIGH-RISK trio (W910 / W911 / W912 + W913 — W922-CONSOLIDATE)

• W909-RESEARCH — 14+ more registry-parity drift candidates surfaced beyond W869's 10. Top three graduated to HIGH-RISK fixes in this batch: cmd_alerts thresholds (W910), _CONFIDENCE_BASES (W911), _DETECTOR_METADATA coverage (W912 + W913). The remaining candidates are queued as W915-W921 drive-by pendings — most are narrow constant-cluster parity gaps the same B+E template will dispatch.
• W910 — HIGH-RISK FIX: cmd_alerts._DEFAULT_THRESHOLDS was missing bottlenecks + dead_exports. Backfilled bottlenecks (>5, WARNING) + dead_exports (>20, INFO) into _DEFAULT_THRESHOLDS. Pinned with a new 3-test parity lint at tests/test_w910_alerts_threshold_parity.py asserting _DEFAULT_THRESHOLDS / _TREND_LABELS / _WORSE_WHEN_* stay in lockstep. 46/46 focused tests green. Missing thresholds meant agents calling roam alerts against either metric got silent-fallback behavior — Pattern 2 silent-fallback at the threshold layer.
• W911 — HIGH-RISK FIX: _CONFIDENCE_BASES derived from canonical findings constants. src/roam/catalog/detectors.py:_CONFIDENCE_BASES now derives from roam.db.findings.CONFIDENCE_HEURISTIC / CONFIDENCE_STRUCTURAL / CONFIDENCE_STATIC_ANALYSIS / CONFIDENCE_RUNTIME instead of carrying its own string literals. Frozenset shape preserved; zero outside callers needed updates. New parity test tests/test_w911_confidence_tier_parity.py (3 tests). 162 focused tests pass. Closes a Pattern 3a vocabulary-divergence path between the detector catalog + the findings registry — confidence-tier strings are now sourced from one canonical site.
• W912 — HIGH-RISK LINT: detector metadata coverage gap pinned. tests/test_w912_detector_metadata_coverage.py (3 tests) asserts every _QUERY_COSTS task_id has a matching _DETECTOR_METADATA row. Pre-fix gap: 11 detectors silently fell back to default precision/impact (broad-except-swallow / async-* / dangerous-eval / etc.); test originally xfail to document the gap.
• W913 — Detector metadata backfill. Backfilled the 11 missing _DETECTOR_METADATA rows in src/roam/catalog/detectors.py:4001-4020 with deliberate per-detector precision/impact picks. xfail removed from W912 lint. Parity is now 34 task_ids ↔ 34 metadata rows. 17/17 focused tests pass. Closes the W912 silent-fallback gap.

Changed — W877/W878/W879/W880 hoist arc carry-through + W894 confidence-tier mismatch sealed (W922-CONSOLIDATE)

• W877 — _enclosing_symbol hoist landed in detail. Defensive variant from smells.py chosen as canonical at src/roam/catalog/_shared.py; type_switch.py's try/except OperationalError contract preserved. 253 focused tests pass.
• W878 — _bare_command_name QUADRUPLE-mirror SEALED. W874 originally identified a triple-mirror; the dispatch surfaced a fourth twin (modes/policy._normalise_command) that the literal-string grep missed. All four sites consolidated into a new module src/roam/commands/_command_utils.py (bare_command_name, 42 LOC). -47 +33 across 3 patched files; 158 focused tests pass. Captured methodologically as drive-by W920: literal-string clone-detection misses semantically-equivalent rename-variants; the W855 behavioral-fingerprint detector could replace the literal sweep.
• W879 / W880 / W901 — Catch-up commit-set notes. Already shipped in the W908 batch; line items re-pinned here so the W922 changelog reads stand-alone. _camel_split canonical at search/index_embeddings.py + __all__ export; _parse_iso canonical at evidence/approval.py + the OLD "duplicated here to avoid import cycle" docstring DELETED (no cycle existed).
• W894 — Inline fix landed. Temporal-coupling confidence-tier mismatch fixed at the hand-rolled site: parent = heuristic (git-cochange frequency); cluster rollup = structural (graph aggregation over heuristic-tier findings). Decorator side aligned. W867 lint extended with a value-parity test so the same drift can't recur silently.

Hardened — W902 cargo-cult follow-through + CLAUDE.md anti-pattern rule (W904 / W905 / W907 — W922-CONSOLIDATE)

• W904 — django_post.py docstring corrected: the alleged duplication NEVER EXISTED. Triple-false claim audited and removed (no cycle + no duplication + confused readers). python_lang.py is Django-agnostic; the docstring's "duplicated from python_lang" hedge was pattern-matched cargo-cult, not a factual claim about the codebase.
• W905 — cmd_oracle.py:83 lazy import PROMOTED to module-level. Companion try/except masking the impossible ImportError REMOVED. The "duplicated here to avoid import cycle" docstring was factually wrong; no cycle exists; the defensive try/except was dead code. 316 focused tests pass.
• W907 — CLAUDE.md "Verify the cycle before hedging" sub-section landed. Added between "Never N/A without running it" and "Adding-a-command checklist" in the Quality-discipline section. Codifies the cargo-cult anti-pattern that W904 + W905 + W880 collectively exposed (3+ false-hedge replications across unrelated modules in the same audit). The rule: before writing "duplicated here to avoid X" in a docstring, actually verify X exists; if X turns out to be false the hedge is a fabricated rationalisation that confuses the next reader.

Hardened — W914 second stale-pending re-triage (W922-CONSOLIDATE)

• W914 — 8 more stale-pending closures + 1 supersession. Continuation of W876 methodology. Flipped pending → completed: W336 / W362 / W370 / W370b / W371 / W383 / W399 (duplicate-pending shadows of already-completed waves). W356 marked superseded as obsolete process directive. Combined with W876, the two passes have flipped 19 stale-pending tasks total — the legacy "Pending after WXXX" sections have visibly contracted across the W886 → W908 → W922 arc.

Operational findings (W922 batch — pending fix / decision)

• W915 — _QUERY_COSTS closed-enum-as-string-literals. Same Pattern 3a vocabulary shape as W911; the keys live as bare string literals rather than canonical-constant references. Trivial hoist behind a _QUERY_COST_KEYS frozenset.
• W916 — CLAUDE.md should cite findings.py as the confidence canonical. Post-W911, the canonical site for confidence-tier strings is roam.db.findings.CONFIDENCE_*. CLAUDE.md's "Confidence-tier vocabulary" sub-section should cite the module + constant names explicitly so future detectors don't reinvent.
• W917 — test_smells_confidence_mapping_parity hardcoded-string set should derive from findings. The W867 lint currently hardcodes its allowed-tier set; should derive from roam.db.findings.CONFIDENCE_* for the same reason W911 flipped _CONFIDENCE_BASES.
• W918 — _resolved_thresholds silent fallback for unknown metrics. Returns a default threshold on unknown metric names. Should raise OR surface a partial_success=True envelope so agents calling roam alerts <new-metric> don't get a silent default-pass.
• W919 — TypedDict for cmd_alerts rule shape. The rule dict carries threshold / direction / severity / trend_label / worse_when_* ad-hoc; a TypedDict would surface drift at write time (the W910 backfill would have failed at type-check time on a TypedDict-annotated _DEFAULT_THRESHOLDS).
• W920 — Differently-named-twin audit via the W855 behavioral-fingerprint detector. W878 surfaced a 4th literal-named twin that grep missed; the W855 rename-invariant clone detector (already shipped) could replace literal-string sweeps for this class of audit. Worth proving on one more case before generalising.
• W921 — Audit other "duplicated from python_lang" claims. W904 follow-up; sweep the codebase for other "duplicated from python_lang" / "mirrors python_lang" hedges and verify each one is factually true. The W904 finding suggests this is a recurring template.
• W903 — W686 path-length recurrence operational note. Recurring across batches; tooling-side investigation not addressable from inside roam.

Added — Canonical-source consolidation arc (W923 / W925 / W929 / W935 + W866 / W920 / W927 / W928 — W939-CONSOLIDATE)

• W923 — REAL clone target: make_smell_finding(...) hoisted to src/roam/catalog/_shared.py. 4 catalog-layer callers migrated. smells.py + type_switch.py consume the canonical via direct import alias; parallel_hierarchy.py + clones_cross_layer.py route dict construction through canonical via detector-specific arg-adapter wrappers. Optional kwargs are OMITTED FROM DICT when None — preserves the 8-key envelope shape every finding-registry test asserts. The W855 rename-invariant detector reports 0 remaining catalog-layer _finding clone pairs. 237 focused tests pass. Closes the W874-finding clone family inside the detector code that catches clones.
• **W935 — make_finding_id(prefix, subject, *raw_parts) hoisted to roam.db.findings. 6 sites (cmd_audit_trail_conformance / cmd_bus_factor / cmd_dead / cmd_doctor / cmd_orphan_imports / cmd_smells) reduced their _XXX_finding_id bodies to one-line returns. All 6 outputs hash-byte-identical before/after — persisted finding rows stay valid. 5 dangling import hashlib lines removed inline. 77/78 focused tests pass** (1 pre-existing unrelated failure).
• W925 — detectors._finding fully annotated matching smells.make_smell_finding style (sqlite3.Row for sym, Mapping[str, Any] / Iterable[str] / int | None for kwargs, -> dict return). 230 focused tests pass. Pairs with W923's canonical hoist to give the catalog-layer finding constructors complete type coverage.
• W929 — _RE_CAMEL_SPLIT + _RE_UPPER_SPLIT canonical at tfidf.py. search/tfidf.py owns the canonical pre-compiled regexes; index_embeddings._camel_split() is now a thin wrapper consuming them. Option (C) chosen after option (A) hit a circular import — captured as the operational pattern: when canonical-hoist hits a cycle, owner-flip is the safe alternative.
• W866 — Dispatch-table refactor on 3 type-switch sites W852 flagged on roam-code's OWN code. smells.py:1793 + smells.py:1812 (magic-numbers walker; 4-arm isinstance chain → _AST_HANDLERS dispatch by type(child)) and registry_dispatch.py:170 (3-arm dispatch on type(value) against ast.Dict / List / Tuple). Dogfood-OCP win: the W852 type-switch detector is now clean against itself. 216 focused tests pass.
• W919 — AlertThreshold TypedDict landed in cmd_alerts.py. Closes the W919 drive-by from the W922 batch. Fields: op as Literal[5 comparators], value as float | int, level as str (deliberately not Literal — _resolved_thresholds normalizes UPPER-case at load time). _DEFAULT_THRESHOLDS typed as dict[str, AlertThreshold]. 49/49 focused tests pass.
• W915 — QUERY_COST_LOW / _MEDIUM / _HIGH constants added to detectors.py with _QUERY_COSTS deriving from them. Closes the W915 drive-by from the W922 batch — the keys no longer live as bare string literals.
• W917 — _SMELL_CONFIDENCE_TIERS (3-of-4 subset, no runtime) added to test_smells_confidence_mapping_parity.py. test_all_confidence_values_are_canonical now uses the smells-specific allowlist instead of the global canonical set. Closes the W917 drive-by from the W922 batch.

Hardened — Behavioral-fingerprint twin sweep + cycle-verification discipline (W920 / W927 / W928 — W939-CONSOLIDATE)

• W920 — Behavioral-fingerprint sweep for differently-named twins (Explore mode). 5 unmigrated twins surfaced beyond literal-grep reach: relations.py:343 _is_test_path (W873 left as cycle-break — W902 method says verify); pytest_fixtures.py:112 _is_test_function; rerank.py:376 inline (4+ call sites); cmd_adversarial.py:347 inline; cmd_next.py:379+518 inline (later DECLASSIFIED as non-clones — shape-checks, not parsers). Methodologically validates W855 rename-invariant detector > literal-string grep for clone-family completeness.
• W927 — rerank.py:376-396 inline 21-line OR-chain extracted to module-level _is_test_path() + 4 named pattern tuples. Did NOT delegate to is_test_file — would broaden behavior (rerank was tuned WITHOUT conftest.py / _test.java / etc). 26-case truth table at .audit-tmp/verify_rerank_helper.py confirms 0 diffs before/after. 139 retrieve tests pass. Cycle-verification discipline (W907) applied: behavior-narrowness preserved by deliberate non-delegation.
• W928 — relations.py:343 cycle verdict NO (AST transitive scan: changed_files imports only file_roles + test_conventions + git_utils, never reaches relations). The W873-era "to avoid roam.commands import cycle" comment was cargo-cult false per the W902 method. BUT delegation would broaden behavior — relations._is_test_path is narrower than the canonical. Kept local def, REPLACED misleading comment with W928's verification record + "deliberately narrower; broadening requires reindex audit" rationale. 31 index tests pass. Closes a W907-rule case: cycle was false but delegation was still wrong for behavior reasons.

Hardened — W930 declassification + W907 carry-through (W939-CONSOLIDATE)

• W930 — Closed not-applicable. cmd_next.py inline startswith("roam ") usages are shape-checks, not parsers (W920 misclassified them as twins). Captured the W920 audit's own false-positive class.
• W916 — CLAUDE.md confidence-tier vocabulary section now cites src/roam/db/findings.py canonical with the 4 CONFIDENCE_* constant names + "extend canonical first, never hardcode at consumer site" discipline rule. Pairs with W911 / W917 to give the confidence-tier vocabulary a single source of truth + reader-facing pointer.

Operational findings (W939 batch — pending fix / decision)

• W931 — Add mypy to .venv typecheck extras. Discovered while running W919 / W925 type-annotation validation; convenience pending. | pyproject.toml [project.optional-dependencies] | 30 min
• W932 — Audit detectors._finding callers for non-dict evidence= (W925 follow-up). Type annotation says Mapping[str, Any] | None; callers should be audited for stray non-dict shapes. | src/roam/catalog/detectors.py callers | 1h
• W933 — Tighten cmd_alerts._parse_alerts_yaml + _resolved_thresholds return types (W919 follow-up). The TypedDict landed; the two YAML-loader return types should now narrow from dict[str, dict] to dict[str, AlertThreshold]. | src/roam/commands/cmd_alerts.py | 1h
• **W934 — test_findings_* parametrization opportunity** (W923 cluster). The 4 catalog-layer migration sites have near-identical test scaffolding; parametrize for drift-guard discipline. | tests/test_findings_*.py | 1-2h
• **W936 — Migrate query_cost string-literal defaults to QUERY_COST_*** (W915 follow-up). Consumer sites that take a query_cost kwarg with a default string literal should now reference the new constants. | grep-then-migrate | 1h
• W937 — Sweep mis-encoded Unicode arrows in docstrings (W929 drive-by). Captured while editing tfidf.py; some docstrings have UTF-8-mangled arrows from prior edits. | grep → / mangled variants | 30 min
• W938 — Fold cmd_bus_factor._repo_summary_finding_id (W935 4th cousin). Has the same shape as the 6 sites already migrated but takes only prefix + subject, no raw_parts. Migration is mechanical once W935 is reviewed. | src/roam/commands/cmd_bus_factor.py | 30 min

W949-CONSOLIDATE — GATE 1 of the registry-parity milestone CLOSED (W940 → W941 + W871-bulk + W895 / W896 / W897 + W870 + W914-pass-3 + W938)

> MILESTONE: this is NOT just another batch. The W869 research memo > catalogued the registry-parity bug class as having 10 instances across > roam-code; Instance #1 (the smell-detector P0 surface) is now > structurally CLOSED. The pre-state had TWO hand-rolled parallel data > tables in lockstep — ALL_DETECTORS = [...] in src/roam/catalog/smells.py > + _SMELL_KIND_TO_CONFIDENCE = {...} in src/roam/commands/cmd_smells.py > — held together by W862 + W867 drift-guard lints (catching the symptom) > after every detector wave. W941 converted BOTH tables to DERIVED VIEWS > off the @detector-decorated registry: every detector self-registers > with its smell_id + confidence tier; the parallel-data shape is now > structurally impossible for this registry. ~78 lines of hand-rolled > parallel data eliminated. 24 detectors + 1 rollup = 25 confidence > entries, all canonically registered, all derivable. Hash-stability > mandate held (detector output bytes unchanged). 283/283 focused tests > pass. Gate 1 of W940's milestone framing is CLOSED; the > registry-parity bug class is no longer maintainable as parallel-data > debt for this surface. The remaining 9 instances (MCP tool registry, > mode-allowlists, _DEPRECATED_COMMANDS, subject_kind, etc.) are > sequenced in W869 + W940 follow-up memos.

Closed — GATE 1 registry-parity milestone (W940-RESEARCH → W871-bulk → W941 — W949-CONSOLIDATE)

• W940-RESEARCH — Registry-parity next-wave sequencing memo. Ranked 10 candidate waves on the W869 backlog and recommended doing the W895 / W896 / W897 design calls + W871-bulk migration FIRST rather than patching individual drift instances one-by-one. The decorator migration is nonlinear: it eliminates a debt class permanently for the registry it covers, vs the per-instance patches which only seal individual cases. This memo is the strategic load-bearing input for the W941 close.
• W895 — @detector rollup_kinds={"cluster": tier} kwarg. Design closure inline: replaces W871's POC register_rollup_kind orphan-API with a kwarg on the @detector decorator. Single source of truth: a detector that emits a rollup smell_id (e.g. temporal-coupling-cluster) declares it inline. Captured as W943: decide formal orphan-API status for the standalone register_rollup_kind helper.
• W896 — all_detectors() returns sorted-by-smell_id. Design closure inline: SARIF emitters + golden-fixture tests depend on stable run-to-run ordering. Sorted retrieval makes the registry grep-friendly and decoration-order-independent.
• W897 — freeze_registry() called at run_all_detectors entry. Design closure inline: validator runs once per invocation (not per @detector call), so import order + decoration order are decoupled. Resolves the W871 POC's open finalisation-semantics question.
• W871-bulk — 22 detectors migrated to @detector decorator. All remaining smells.py detectors now self-register via the decorator + register_rollup_kind (now expressed via the rollup_kinds= kwarg per W895). registry.py extended with W895/W896/W897 implementations inline. temporal-coupling cluster rollup upgraded to rollup_kinds={"cluster": CONFIDENCE_STRUCTURAL} per W895. W862 + W867 parity lints flipped SUBSET → EQUAL (the registry IS now the source of truth, not a subset of the hand-rolled table). 283/283 focused tests pass.
• W941 — THE GATE 1 CLOSURE: ALL_DETECTORS + _SMELL_KIND_TO_CONFIDENCE converted to DERIVED VIEWS. src/roam/catalog/smells.py:ALL_DETECTORS is now [d.fn for d in registry.all_detectors()] (sorted-by-smell_id per W896); src/roam/commands/cmd_smells.py:_SMELL_KIND_TO_CONFIDENCE is now {d.smell_id: d.confidence_tier for d in registry.all_detectors()} | {rollup_id: tier for ...} per W895 rollup_kinds. ~78 lines of parallel-maintained data eliminated; 24 detectors + 1 rollup = 25 confidence entries, all canonically registered, all derivable. Detector output bytes unchanged (hash-stability mandate held). 283/283 focused tests pass. The W869-catalogued registry-parity bug class is now structurally impossible for this registry surface.
• W870 — Per-detector version-stamp parity lint (permissive). New AST lint asserts that every @detector-registered detector either (a) has a per-id <DETECTOR>_VERSION module constant OR (b) inherits the composite SMELLS_DETECTOR_VERSION fallback. 7/24 detectors have per-id stamps; 17/24 share the composite. Captured as W944 for a strict-mode toggle once per-id stamps cover the full roster. 3/3 lint tests pass. Final P0 piece of W869's hybrid Archetype B+E recommendation.

Closed — Hash-stable canonical-source fold (W938 — W949-CONSOLIDATE)

• W938 — cmd_bus_factor._repo_summary_finding_id folded onto W935's make_finding_id canonical. 4th-cousin site captured in the W939 batch (no *raw_parts, only prefix + subject). One-line return; hash-stable across 5 sample inputs (verified byte-identical pre/post). import hashlib removed. 43 focused tests pass. W935 finding-id-builder family is now fully consolidated across all 7 sites; no remaining outliers.

Closed — Third stale-pending triage (W914-pass-3 — W949-CONSOLIDATE)

• W914-pass-3 — Third pass over the stale-pending roster. 3 confirmed closures (W221 user-blocked + sub-scope absorbed into W196 / W199 / W202 / W203 milestone; W354 verification absorbed into W454 qualified_only flag; W367 duplicate-pending already completed as #513), 2 BLOCKED (W350 / W351), 11 STILL VALID. Combined with W876 + W914: 22 stale-pending tasks flipped across 3 passes. Pattern locked: triage in waves, document the closure reason, never silently drop.
• W937 — Closed not-applicable. Source sweep confirmed no β†’ Unicode-arrow corruption remains; W929 fix was the only instance.

Operational findings (W949 batch — pending fix / decision)

• W942 — Pivot W862 count-drift lint to registry source. W862 currently asserts len(ALL_DETECTORS) == <docstring count>; after W941 converted ALL_DETECTORS to a derived view, the docstring-count assertion is structurally redundant. Pivot to len(registry.all_detectors()) == <docstring count> (registry-side authority). Captured as DBD-3 in the W940 sequencing memo. | tests/test_smells_detector_count_drift.py | 30 min
• W943 — Decide register_rollup_kind orphan-API status. W895 added rollup_kinds= kwarg on @detector; the standalone register_rollup_kind helper from the W871 POC is now functionally redundant. Decide: deprecate-and-remove vs leave-as-explicit-API. Captured as DBD-1 in the W940 memo. | src/roam/catalog/registry.py | 30 min decision + 1h migration
• W944 — W870 strict-mode toggle for per-detector versioning. Once per-id <DETECTOR>_VERSION stamps cover the full 24-detector roster (currently 7/24 have them), flip the W870 lint from permissive (composite fallback OK) to strict (per-id required). | src/roam/catalog/registry.py + per-detector module constants | 2-4h
• W945 — Refresh registry.py docstring "two SOURCE-OF-TRUTH" comment (W941 follow-up). The module docstring still implies two parallel sources; after W941 the registry IS the single source. Cosmetic. | src/roam/catalog/registry.py docstring | 30 min
• W946 — Refresh smells.py:19 parallel_hierarchy wording (W941 follow-up). Comment narrates the pre-W941 two-table arrangement. | src/roam/catalog/smells.py:19 | 15 min
• W947 — Simplify test_decorator_registry_parity self-referential assertions (W941 follow-up). Post-W941 the parity lint compares a derived view against its own source — partially tautological. Replace with a value-shape contract test instead of identity. | tests/test_smells_confidence_mapping_parity.py | 1h
• W948 — Move tier rationale inline to @detector calls (W941 follow-up, medium effort). Per-detector confidence-tier rationale currently lives in CLAUDE.md prose; with W895's rollup_kinds= kwarg the per-detector tier is already declared at the decorator. Adding a short inline # heuristic — name pattern only comment at each @detector(...) call would surface the rationale where the next reader will look. | smells.py + sibling catalog modules | 2-3h

W965-CONSOLIDATE — Gate-1 cleanup + W525 strategic pause + W918 / W924 / W933 typing/silent-fallback close

> CONSOLIDATE checkpoint = W965. ~10 closures + 15 drive-by captures > since W949-CONSOLIDATE. Two arcs landed in parallel: (1) the W942 / W945 > / W946 / W947 / W955 / W956 follow-throughs that close the W941 Gate-1 > cleanup queue (count-drift lint pivoted to registry source; pre-W941 > "two SOURCE-OF-TRUTH" wording flipped to past-tense across registry.py > + smells.py + the parity test; freeze_registry invariant numbering > re-ordered to match execution order), and (2) three independent > source-tightening landings — W918 closing a Pattern 2 > silent-fallback hole in _resolved_thresholds (unknown user-supplied > metric now surfaces via warnings_out + envelope partial_success=True > + a new agent_contract.facts entry), W924 stamping > detector_version on every detectors._finding envelope via the > pre-existing canonical roam.catalog.versions.detector_version(task_id) > (most task_ids → DEFAULT_VERSION='1.0.0'; the nested-lookup site > carries the 1.1.0 override), and W933 tightening > cmd_alerts._parse_alerts_yaml to dict[str, dict[str, Any]] + > selecting Option B (loose-but-honest typing) on _resolved_thresholds > because slot.update(rule) precludes TypedDict without runtime > validation. W525 — STOP AT INVENTORY. The W869 Instance #2 proving > ground (MCP tool registry) ran the inventory pass and surfaced real > structural gaps the wave's prescribed derivation would have silently > papered over: the hand-rolled _CORE_TOOLS (57 tools) does NOT match > @roam_capability(category="core") (0 — the category doesn't exist on > the decorator), and mcp_preset=("core",) is mostly boilerplate (228 > of 230 tools carry it). The decision was made to STOP at inventory > rather than mechanically derive until the strategic call on > derivation source (category= vs mcp_preset= vs hand-rolled) lands as > W357 long-horizon work. W525 split into 5 deep drive-bys > (W950 strategic / W951-W953 evidence / W954 closed) feeding that call. > W954 regression-guard test landed — tests/test_w954_core_tools_capability_drift.py > (3 tests, 191 lines, all pass) snapshots _CORE_TOOLS=57, capability > registry=230 (one retired), mcp_preset="core" boilerplate=228, > category="core"=0, and floors at ~10% headroom (≥18 in_core_not_cap, > ≥180 in_cap_not_core). Hash-stability mandate held trivially across > the batch — W924's stamp lands on the dict AFTER make_finding_id > hashes only *raw_parts. Stale-pending triage closed three more rows > via the W914 methodology (W221 / W354 / W367 confirmed-closed from the > W949 batch; carried into the W965 BACKLOG strike-throughs for > visibility).

Closed — Gate-1 cleanup follow-throughs (W942 / W945 / W946 / W947 / W955 / W956 — W965-CONSOLIDATE)

• W942 — Count-drift lint pivoted to registry source. tests/test_smells_detector_count_drift.py no longer reads len(ALL_DETECTORS) (post-W941 a derived view — the docstring assertion was structurally self-referential). All 5 call-sites updated to call len(list(all_detectors())) from roam.catalog.registry. 179 focused tests pass. Closes the DBD-3 follow-up from the W940 sequencing memo.
• W945 — registry.py docstring + comments refreshed to past-tense. Lines ~11-16: "two SOURCE-OF-TRUTH dicts" framing now narrates the pre-W941 arrangement in past tense. Lines ~68-72: "The source-of-truth collections..." replaces the old present-tense phrasing. Cosmetic; closes the W941 follow-up so the next reader doesn't believe two parallel sources still exist.
• W946 — smells.py module docstring refreshed. Lines 12-20: notes ALL_DETECTORS is now a derived view; registration happens via @detector decorator + detector(...)(fn) calls. Removes the lingering pre-W941 "two parallel hierarchies" claim.
• W947 — Regression-guard note pinned in test_decorator_registry_parity.py. 11-line "W947 note (KEPT as regression guard, do not delete)" block added at the top of the file. Captures why the parity lint stays even though post-W941 it compares a derived view against its own source — the lint is the regression guard against silent un-deriving (e.g. a future refactor that hand-rolls a row back into _SMELL_KIND_TO_CONFIDENCE).
• W955 — Inline tighten of pre-W941 transition wording. tests/test_decorator_registry_parity.py:9 flipped "belt-and-braces during the transition window" → past-tense ("…during the transition window from W871 → W941"). Single-line tweak; closes the W947 audit's drive-by.
• W956 — freeze_registry invariant numbering re-ordered to match execution order. The docstring numbered the three invariants 1/2/3 in a different order than the code body actually checked them. Re-numbered to match execution order: (1) duplicates first (cheapest), (2) anchored ids, (3) canonical tier. Closes a subtle reader-trip-up that would have wasted the next debugger's time.

Closed — Source-tightening trio (W918 / W924 / W933 — W965-CONSOLIDATE)

• W918 — Pattern 2 silent-fallback fix in _resolved_thresholds. src/roam/commands/cmd_alerts.py _resolved_thresholds previously returned a default threshold (op='>', value=0) on unknown user-supplied metrics. Fix: new warnings_out: list[str] | None = None parameter accumulates a per-unknown-metric warning string; envelope flips summary.partial_success = True whenever any warning was raised; new agent_contract.facts entry surfaces "user-supplied metric <name> was not in the canonical roster — defaulted to op='>', value=0". Backward compat preserved (existing .roam/alerts.yaml configs untouched). 52 focused tests pass. Closes the W918 W922-carry-forward; pairs with the W910 backfill + W911 canonical-derivation work to give the alerts surface a complete Pattern 2 + Pattern 3a discipline.
• W924 — detector_version stamp on detectors._finding. src/roam/catalog/detectors.py:_finding now stamps finding["detector_version"] via the pre-existing canonical roam.catalog.versions.detector_version(task_id). Most task_ids resolve to DEFAULT_VERSION='1.0.0'; the nested-lookup site carries the 1.1.0 override per the W81 ABC contract. Hash-stability verified — make_finding_id hashes only *raw_parts (the W935 contract); the stamp lands on the dict AFTER id computation, so persisted finding rows stay byte-identical. 219 focused tests green. Closes the W924 carry-forward from the W922 batch; pairs with the W81 per-component version-stamp arc.
• W933 — cmd_alerts return-type tightening. _parse_alerts_yaml flipped to dict[str, dict[str, Any]] (previously dict[str, dict]). _resolved_thresholds picked Option B (loose-but-honest): keeping the return type as dict[str, dict[str, Any]] rather than dict[str, AlertThreshold] — the closer TypedDict was rejected because the body does slot.update(rule) and TypedDict mutation paths require runtime validation that would bloat the hot path. The W919 AlertThreshold TypedDict is still in place for the _DEFAULT_THRESHOLDS shape contract; the looser return type accommodates the user-supplied-rule merge path. 46/46 focused tests pass. Closes the W933 W939-carry-forward.

Captured — W525 STOP-AT-INVENTORY decision + W954 regression-guard (W950 / W951 / W952 / W953 / W954 — W965-CONSOLIDATE)

• W525 — W869 Instance #2 proving ground: STOP AT INVENTORY. The Gate-2 candidate (MCP tool registry; 224 wrappers / 57 in _CORE_TOOLS) ran the inventory pass and surfaced real structural gaps the W869 hybrid Archetype B+E template would have silently papered over: (a) hand-rolled _CORE_TOOLS (57) does NOT match @roam_capability(category="core") because the category enum doesn't include "core" at all (0 hits) — category= is shaped around code-intelligence / governance / etc., not preset membership; (b) mcp_preset=("core",) is the closest derivation source but it's mostly boilerplate — 228 of 230 tools carry it as a copy-paste default, not a curated subset. Decision: STOP at inventory until the strategic call on derivation source (category= vs mcp_preset= vs hand-rolled) lands as W357 long-horizon work. The W869 wave's prescribed derivation was unsafe on this surface.
• W954 — Regression-guard test landed. New tests/test_w954_core_tools_capability_drift.py (3 tests, 191 lines, all pass). Snapshot: _CORE_TOOLS=57, capability registry=230 (was 231 — one retired since W949), mcp_preset="core" boilerplate=228, category="core"=0. Floors at ~10% headroom (≥18 in_core_not_cap, ≥180 in_cap_not_core) so the test doesn't false-trip on additive drift but catches structural collapses (e.g. a future wave silently shrinking _CORE_TOOLS to align with mcp_preset="core"). Closes the W525 drive-by chain inline.
• W950 / W951 / W952 / W953 — Deep drive-bys captured. W950 STRATEGIC: pick category= vs mcp_preset= path for MCP registry derivation (feeds W357). W951: mcp_preset default ("core",) is dead metadata — most call-sites accept it via copy-paste rather than deliberate curation. W952: 24 MCP-only tools have no @roam_capability anchor at all (gap class to address before any derivation pass). W953: 4 naming-drift cases between CLI command names + MCP wrapper names that the W869 template would not catch (e.g. roam_pr_replay vs roam pr-replay). All four captured as pendings for the W965+ wave to triage as the strategic answer lands.

Closed — Stale-pending triage (W221 / W354 / W367 — W965-CONSOLIDATE carry-from-W949)

• W221 — User-blocked + sub-scope absorbed. Audit Trail (R29) snapshot work absorbed into the W196 / W199 / W202 / W203 producer-wiring + milestone-integration arc; no remaining net-new scope. Closed via W914-pass-3.
• W354 — Verification absorbed into W454. qualified_only flag verification work absorbed into the W454 implementation; no separate verification scope remains. Closed via W914-pass-3.
• W367 — Duplicate-pending. Already completed as #513 in a prior session; row was stale. Closed via W914-pass-3.

Operational findings (W965 batch — pending fix / decision)

• W357 (strategic, long-horizon) — Pick the MCP registry derivation source. W525 inventory pass surfaced three candidates (@roam_capability(category=...), mcp_preset=(...), hand-rolled _CORE_TOOLS) — each has structural gaps. Strategic decision required before any derivation pass on the MCP tool registry. | src/roam/mcp_server.py + src/roam/plugins/capability.py | TBD (strategic)
• W950 — STRATEGIC: pick category= vs mcp_preset= path for MCP registry derivation. Sub-question of W357. Feeds the W869 Instance #2 wave once it unblocks. | strategic | TBD
• W951 — mcp_preset=("core",) default is dead metadata. 228 of 230 tools carry the default value via copy-paste, not curation; a derivation pass that trusts the metadata would over-include. Decide: strip the default, or curate it deliberately. | src/roam/mcp_server.py decorator default | 1-2h decision + 4-6h migration
• W952 — 24 MCP-only tools have no @roam_capability anchor. Gap class to close before any category=-based derivation pass. | per-tool audit | 4-6h
• W953 — 4 naming-drift cases between CLI + MCP wrappers. Tools where the MCP wrapper name does not derive from the CLI command name via the canonical kebab-→-snake transform. Captured for a documentation-grade audit before W357 lands. | per-case docs | 2h
• W957 — W862 lint "Fix:" hint forward-compat nit. Post-W942 pivot, the lint's "Fix: update the docstring" hint references the registry rather than ALL_DETECTORS; nit-pick wording polish. | tests/test_smells_detector_count_drift.py hint string | 15 min
• W958 — _load_alerts_config return-type tightening. Companion to W933 — the YAML config loader returns dict[str, dict] ad-hoc; tighten to dict[str, dict[str, Any]] for consistency with _parse_alerts_yaml. | src/roam/commands/cmd_alerts.py | 30 min
• W959 — _check_thresholds Alert TypedDict bundle. Companion to W933 — the per-finding Alert dict shape would benefit from a TypedDict declaration analogous to AlertThreshold. | src/roam/commands/cmd_alerts.py | 1-2h
• W961 — Document uniform naming-drift convention (W954-class follow-up). 15 additional cases beyond W953's 4 surfaced during the W954 audit; a uniform convention doc would deter recurrence. | CLAUDE.md "Adding a new CLI command" section + audit | 2-3h
• W962 — _parse_alerts_yaml op-vocabulary validation at parse time. Pattern 2 family follow-up to W918 — at parse time, validate the comparator is in the closed enum (> / >= / < / <= / ==). Currently the unknown comparator silently falls through to the default. | src/roam/commands/cmd_alerts.py | 1h
• W963 — _check_thresholds unknown-comparator silent skip. Pattern 2 family follow-up to W918 — at check time (not parse time), the unknown comparator silently skips the row instead of surfacing via partial_success. Different code path from W962. | src/roam/commands/cmd_alerts.py | 1h
• W964 — delta_alerts bool coercion silent disable. Pattern 2 family follow-up to W918 — the delta_alerts flag coerces non-bool YAML values via bool(...) rather than surfacing a parse warning; a typo ("yes", "no") silently flips to enabled. | src/roam/commands/cmd_alerts.py | 30 min

W977-CONSOLIDATE — cmd_alerts Pattern-2 family FULLY CLOSED + W923 test-layer consolidation + W966 audit pass

> CONSOLIDATE checkpoint = W977. ~10 closures + drive-by captures > since W965-CONSOLIDATE. Headline: cmd_alerts.py Pattern-2 family is > now FULLY CLOSED end-to-end via the W962 / W963 / W964 trifecta > (op-vocabulary validation at parse + check time; bool-coercion fix on > delta_alerts) followed by the W967 / W968 / W969 trifecta (REAL > BUG: tiny YAML parser silently disabled delta_alerts for users > without PyYAML — a scalar-vs-section detection gap; REAL BUG: > level: "fatal" would KeyError downstream — _CANONICAL_LEVELS > frozenset + _coerce_level helper at 3 sites + counts initializer > fold; drift-guard test pins _VALID_OPS == AlertThreshold.op Literal > via typing.get_type_hints). 87 focused tests pass. Two bugs > were latent (0 fixtures exercised them). This is the SECOND > consecutive Pattern-2 family fully closed in this session — the > first was W826 / W834 / W836 in early-session (silent SAFE on empty > corpus across taint / health / doctor). cmd_alerts.py now > exemplifies the W918 discipline: every silent-fallback path surfaces > via warnings_out + partial_success=true + > agent_contract.facts. W923 test-layer consolidation — > W934 delegated 24 test_<detector>_findings_visible_via_cmd_findings_count > tests to tests/_findings_helpers.py via Strategy C (shared helper, > per-detector tests retained for fixture independence); doctor's > exact-count + critique's tolerant exit-code preserved; 24/24 + > ~190 sibling tests pass; net -46 lines (-114 of actual code). > W966 audit pass — W971 confirmed the codebase was already > W966-compliant: 13 HONEST sites, 0 LYING, 2 VALIDATED. The > "don't TypedDict a boundary you don't validate" discipline existed > before W966 codified it; W933 _resolved_thresholds is the > EXEMPLAR. W975 / W976 added lock-comments at json_envelope + > _compat_profile_payload per W971's recommendations, documenting > the W966 discipline at the call site. Hash-stability mandate held > trivially across the batch.

Closed — cmd_alerts.py Pattern-2 family FULLY CLOSED (W962 / W963 / W964 + W967 / W968 / W969 — W977-CONSOLIDATE)

• W962 — _parse_alerts_yaml op-vocabulary validation at parse time. Pattern 2 family follow-up to W918. Added _VALID_OPS frozenset (> / >= / < / <= / ==); parse-time validation rejects unknown comparators via warnings_out accumulator + partial_success=true. Unknown op now surfaces explicitly instead of silently falling through to default. Warning text follows LAW 2 (imperative) + LAW 4 (concrete-noun terminal). 15 new tests.
• W963 — _check_thresholds unknown-comparator silent skip closed. Pattern 2 family follow-up to W918. Check-time validation now folds through the same _VALID_OPS frozenset; unknown op at check time surfaces via partial_success instead of silently skipping the row. Different code path from W962. Closed inline alongside W962.
• W964 — delta_alerts bool coercion silent disable closed. Pattern 2 family follow-up to W918. New _coerce_bool helper rejects non-bool YAML values (typo "yes" / "no" / 1 / etc.) via warnings_out + partial_success=true instead of silently bool(...)-coercing to enabled.
• W967 — REAL BUG: tiny YAML parser silently disabled delta_alerts for users without PyYAML. New _coerce_scalar helper + scalar-vs-section detection: the fallback parser (when pyyaml is not installed) was treating delta_alerts: true at root as a section header rather than a top-level scalar, silently disabling the feature for the no-PyYAML install path. Fixed by recognising the scalar-vs-section ambiguity at the parser level. 0 fixtures exercised this path pre-fix — confirmed latent.
• W968 — Drift-guard test pinning _VALID_OPS == AlertThreshold.op Literal. New test consumes typing.get_type_hints(AlertThreshold) to extract the op field's Literal[...] members and asserts equality with _VALID_OPS. The next refactor that adds a comparator to the TypedDict but forgets _VALID_OPS (or vice versa) now fails at lint time.
• W969 — REAL BUG: level: "fatal" would KeyError downstream. Added _CANONICAL_LEVELS frozenset + _coerce_level helper at 3 sites + counts initializer fold. Pre-fix: a user-supplied level: "fatal" would parse cleanly into AlertThreshold (the level field is str, not Literal) but downstream code keyed on _LEVEL_ORDER would KeyError. 0 fixtures exercised this path pre-fix — confirmed latent.

Closed — W923 test-layer consolidation (W934 — W977-CONSOLIDATE)

• W934 — 24 test_<detector>_findings_visible_via_cmd_findings_count tests delegated to tests/_findings_helpers.py. Strategy C: shared helper consumes the per-detector fixtures; per-detector tests retained for fixture independence. Doctor's exact-count + critique's tolerant exit-code preserved. 24/24 + ~190 sibling tests pass. Net -46 lines (-114 of actual code), +68 lines of shared helper scaffolding. Closes the W923-cluster test-layer follow-up first surfaced in the W939 batch.

Closed — Typing audit + lock-comments (W958 / W961 / W966 / W971 / W975 / W976 — W977-CONSOLIDATE)

• W958 — _load_alerts_config return-type tightened to dict[str, dict[str, Any]]. Companion to W933 — the YAML config loader's return type now matches _parse_alerts_yaml for consistency.
• W961 — CLAUDE.md MCP tool naming convention section added. New sub-section at CLAUDE.md lines 822-832: documents the uniform roam_<underscored> ↔ <dashed> convention + 4-entry alias allowlist for genuine renames. Closes the W953 / W954 audit's documentation gap.
• W966 — CLAUDE.md "Don't TypedDict a boundary you don't validate" discipline rule. New sub-section at CLAUDE.md lines 156-170 (companion to W907 "Verify the cycle before hedging"). Codifies the W933 Option B decision rationale: TypedDict mutation paths require runtime validation when the body does slot.update(rule) — discipline rule says either validate at the boundary or use the looser dict[str, dict[str, Any]] return type.
• W971 — Audit pass: codebase already W966-compliant. Surveyed 15 TypedDict / boundary sites; 13 HONEST (canonical pattern), 0 LYING, 2 VALIDATED at the boundary. The discipline existed before W966 codified it; W933 _resolved_thresholds is the EXEMPLAR — the Option B return-type widening was the right call structurally.
• W975 — Lock-comment added at json_envelope per W971's audit recommendations. Documents the W966 discipline at the call site for future readers — the summary field is typed as Mapping[str, Any] deliberately because callers do partial-mutation; a TypedDict would force runtime validation downstream.
• W976 — Lock-comment added at _compat_profile_payload per W971's audit recommendations. Same shape as W975 — documents the W966 discipline inline so the next maintainer doesn't tighten the type and trigger a slot.update-class regression.

Operational findings (W977 batch — pending fix / decision)

• W972 — _load_alerts_config non-dict YAML root silent fallback. Real-but-edge bug: a YAML file whose root is a list rather than a dict silently falls through to defaults rather than surfacing via partial_success. Same shape as W918 / W963 / W964. | src/roam/commands/cmd_alerts.py | 1h
• W973 — _make_alert level validation defense. Latent risk: _make_alert doesn't re-validate level against _CANONICAL_LEVELS even though _coerce_level does at construction time. Defense-in-depth tightening. | src/roam/commands/cmd_alerts.py | 30 min
• W974 — Tighten AlertThreshold.level to Literal[...] (now safe per W969). Pre-W969 the field was deliberately str because there was no runtime validation; post-W969 _coerce_level validates at construction time, so the TypedDict Literal tightening is now safe. | src/roam/commands/cmd_alerts.py | 30 min
• W978 — Pre-existing test_bus_factor_stale_kind_emitted failure. Surfaced during W934 consolidation — the test was already failing pre-batch; W934 helper did not regress it but the failure is unrelated. Triage required. | tests/test_findings_bus_factor.py | 1-2h triage
• W979 — dark_matter ↔ dark-matter + fan_symbol Pattern-3a divergence. Surfaced during W934 consolidation — two detector_id slugs use underscores while the rest use kebab-case. Pattern-3a metric-name canonicalisation gap. | src/roam/db/findings.py + per-detector emitters | 1-2h

W1001-CONSOLIDATE — Pattern-2 playbook propagation across 3 candidate modules + SQL ESCAPE discipline + smells_suppress YAML hardening

> CONSOLIDATE checkpoint = W1001. ~15 closures + drive-by captures > since W977-CONSOLIDATE. Headline: the W983-RESEARCH playbook > (synthesised from W977's full cmd_alerts.py Pattern-2 close) was > propagated to the three nominated candidate modules — and the > outcomes are LOAD-BEARING. W987 sealed cmd_smells.py via a > full playbook apply (closed-set --kind validation against > kind_to_confidence() + warnings_out plumbed from the suppression > loader through the CLI to the envelope). W988 was CORRECTLY CLOSED > AS NOT-APPLICABLE — the agent verified W983's premise didn't match > cmd_conventions.py (no user-supplied boundary exists) and STOPPED > instead of fabricating work. W989 sealed cmd_pr_risk.py via a > DIFFERENT real Pattern-2 gap than W983's framing assumed — > _normalise_pr_risk_level was silently flooring unknown input to > "low" per the W718 CI-safety contract; now warns + preserves the > floor; NO TypedDict added per W966 discipline (internal dict, not > user boundary). The methodology lesson: premise verification is > the FIRST step of every playbook application. A playbook that > applies mechanically without checking the premise produces fabricated > work; a playbook that applies with discipline either seals the > nominated gap, finds a different real gap, or stops cleanly. > W990 → W991 SQL ESCAPE sweep: found 10 accidental wildcard sites > + 2 already-correctly-escaped in src/roam/catalog/detectors.py; 3 > were HIGH-risk in the idiom matchers. W991 fixed 8 W990 sites + 6 > parallel-pattern drive-bys + 1 duplicate matmul fallback = 15 LIKE > escapes total; 109 focused tests pass; smoke confirms > finXinXsortedXarray is correctly excluded. W994 + W995 > smells_suppress YAML hardening: W994 found a REAL BUG — > _is_expired was silently defaulting to not_expired on > unparseable expires strings (typo 2026-13-99 → suppression > stayed active). Fix at load-time AND match-time with a new > EXPIRES_FMT constant; 8 tests. W995 surfaced malformed-entry drops > that were previously "silently skipped" per an admitted comment in > the parser — now partitioned into valid/dropped + indexed warnings > + rollup; 7 tests. W982 cmd_fan rename completed: > fan_symbol → fan-symbol (9 cmd_fan.py + ~32 test sites; SQL LIKE > 'fan_%' fixed; the Strategy A persisted-hash break is documented). > 27 focused tests pass. W978 bus_factor stale-kind test fixed via > a fixture monkeypatch — the W405 shallow-history drop of a 2-year-old > commit had made the test brittle; 18/18 pass; 3 sharp drive-bys > captured (W984 autouse conftest / W985 INFO log on the drop / W986 > "first hypothesis" checklist in CLAUDE.md). W983-RESEARCH memo > shipped (dev/CMD-ALERTS-PATTERN-2-CASE-STUDY-2026-05-15.md, > <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> > 374 lines, 7 reusable patterns, 3 candidate modules) and W999 > amended it with the W988/W989 outcomes — the case-study now > explicitly codifies "premise verification is the first step of > every playbook application". W970 CLAUDE.md drive-by: > _DEFAULT_THRESHOLDS is the canonical positive counter-example > for "when TypedDict IS appropriate" — a 7-line paragraph added to > the W966 sub-section pairs the rule with its exemplar. W936 > migrated 37 query_cost string literals to QUERY_COST_* > constants in detectors.py (W939-carry-forward closed). Hash- > stability mandate held trivially across the batch — every fix > either added new validation paths (no pre-fix envelope to compare), > swept LIKE-clause inputs (no detector output bytes moved), or > tightened YAML load-time validation (no persisted finding rows > touched).

Closed — Pattern-2 playbook propagation across 3 candidate modules (W987 / W988 / W989 — W1001-CONSOLIDATE)

• W987 — cmd_smells.py Pattern-2 playbook FULLY APPLIED. Closed-set --kind validation against the canonical kind_to_confidence() set (lazy-derive from registry, not Literal — smart 1-anchor design that survives detector additions); warnings_out plumbed from the suppression loader → CLI → envelope; unknown --kind arguments now surface via warnings_out + partial_success=true + agent_contract.facts entry, matching the W918 / W962 envelope shape. 185 tests pass. Closes the W923-cluster carry-forward Pattern-2 gap.
• W988 — CORRECTLY CLOSED AS NOT-APPLICABLE. The W983 playbook nominated cmd_conventions.py as a candidate but a premise-verification pass found the command has no user-supplied boundary (no --kind / --metric style argument, no YAML-config user input) — the Pattern-2 gap the playbook is designed to seal doesn't exist on this surface. Agent refused to fabricate work; 71 baseline tests still pass (no source bytes moved). DISCIPLINE WIN — the playbook applied with premise-verification rather than mechanically.
• W989 — cmd_pr_risk.py sealed via a DIFFERENT real Pattern-2 gap than W983's framing assumed. The playbook framed the candidate as a slot.update() shape; the actual gap was in _normalise_pr_risk_level — silently flooring unknown input (e.g. level: "foobar") to "low" per the W718 CI-safety contract. Fix: now emits a warnings_out entry + partial_success=true while PRESERVING the W718 floor (CI-safety contract held — unknown levels still fail-closed at the lowest severity, no behavioral regression). NO TypedDict added per W966 discipline: the internal dict is not a user boundary, so the Option B (loose-but-honest) typing is correct. 51 tests pass. Methodologically the most important close in the batch: the playbook propagation found a real bug that wasn't the one the playbook was designed to find — proves the framework gates are working when each application is run with discipline.

Closed — SQL ESCAPE discipline (W990 / W991 — W1001-CONSOLIDATE)

• W990 — SQL LIKE wildcard audit on src/roam/catalog/detectors.py. Found 10 accidental wildcard sites + 2 already-correctly-escaped. 3 sites were HIGH-risk in the idiom matchers (substring matches on symbol names could match across word boundaries with the wildcard interpretation; e.g. find_in_sorted_array would match find%in%sorted%array against finXinXsortedXarray).
• W991 — Fixed 8 W990 sites + 6 parallel-pattern drive-bys + 1 duplicate matmul fallback = 15 LIKE escapes total. Each call site now uses the canonical ESCAPE pattern: LIKE ? ESCAPE '\\' with the parameter pre-escaped via value.replace('\\', '\\\\').replace('%', '\\%').replace('_', '\\_'). 109 focused tests pass. Smoke test confirms finXinXsortedXarray is correctly excluded from the find_in_sorted_array match set post-fix.

Closed — smells_suppress YAML hardening (W994 / W995 — W1001-CONSOLIDATE)

• W994 — REAL BUG: smells_suppress._is_expired silently defaulted to not_expired on unparseable expires strings. A typo like expires: 2026-13-99 (invalid month) was silently treated as "not yet expired", keeping the suppression active forever. Fix: parse the expires value at LOAD time (raises ValueError on unparseable strings via the new EXPIRES_FMT = "%Y-%m-%d" module constant) AND re-validate at match time (defense-in-depth — the match-time path now surfaces a warnings_out entry on unparseable input rather than silently defaulting). 8 tests. The full Pattern-2 envelope shape (warnings_out + partial_success=true) flows through to the suppression loader's caller envelope.
• W995 — Malformed-entry drops now surface (W994 follow-up). Pre-fix the suppression parser had an admitted # silently skipped comment in its malformed-entry handling — invalid YAML entries were dropped without notice. Fix: parser now partitions input into valid + dropped lists; each dropped entry produces an indexed warning string; a rollup count ("dropped N malformed entries (indices X / Y / Z)") accumulates into warnings_out and flips partial_success=true. 7 tests. Same envelope shape as W994.

Closed — Source carry-forwards + drive-bys (W936 / W970 / W978 / W982 — W1001-CONSOLIDATE)

• W982 — fan_symbol → fan-symbol rename completed. 9 source sites in src/roam/commands/cmd_fan.py + ~32 test sites updated; the SQL LIKE 'fan_%' pattern fixed (Pattern-3a kebab-case canonicalisation, sister to W979's dark_matter/fan_symbol divergence). Strategy A persisted-hash break documented: existing findings_history rows with subject_id LIKE 'fan_symbol:%' will not match the new 'fan-symbol:%' pattern until the next reindex; the migration is forward-only by design. 27 focused tests pass.
• W978 — test_bus_factor_stale_kind_emitted failure FIXED via fixture monkeypatch. Root cause: W405's shallow-history truncation was dropping a 2-year-old commit the test fixture depended on. Fix: fixture monkeypatches the cutoff_days to preserve the test's expected stale-kind flow. 18/18 tests pass. Closes the W934 drive-by (W978 carry-forward). Three sharp drive-bys captured during the fix (W984 autouse conftest / W985 INFO log on the drop / W986 "first hypothesis" checklist in CLAUDE.md).
• **W936 — Migrated 37 query_cost string literals to QUERY_COST_* constants in src/roam/catalog/detectors.py.** Pairs with W915's QUERY_COST_LOW/MEDIUM/HIGH constant introduction (W939-CONSOLIDATE) — the consumer sites now reference the canonical constants instead of bare strings. Closes the W939 carry-forward.
• W970 — CLAUDE.md W966 sub-section gained a 7-line positive counter-example. Added a paragraph naming _DEFAULT_THRESHOLDS (in cmd_alerts.py) as the canonical "when TypedDict IS appropriate" exemplar — the dict is constructed from a closed shape, no slot.update() mutation path exists, and the TypedDict surface (W919 AlertThreshold) accurately captures it. Pairs the W966 rule ("don't TypedDict a boundary you don't validate") with its inverse exemplar so the next reader sees both sides of the discipline.

Research + memos (W983-RESEARCH / W999 — W1001-CONSOLIDATE)

• W983-RESEARCH — dev/CMD-ALERTS-PATTERN-2-CASE-STUDY-2026-05-15.md (374 lines, 7 reusable patterns, 3 candidate modules). Synthesises the full W977 cmd_alerts.py Pattern-2 close into a reusable playbook: (1) warnings_out: list[str] | None accumulator on every boundary parser/normaliser; (2) _VALID_ closed-set frozensets at the boundary; (3) parse-time + check-time validation paths share the same frozenset; (4) latent-bug discipline (0-fixture paths are silent bugs); (5) typing.get_type_hints drift-guard pinning TypedDict Literal[...] ↔ _VALID_ frozenset; (6) coerce-helper pattern at construction time; (7) _CANONICAL_* frozenset for vocabulary that downstream code dictionaries-keys on. Three nominated candidates: cmd_smells (W987), cmd_conventions (W988), cmd_pr_risk (W989). <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• W999 — W983 case-study amendment. Added a new "Premise verification is the first step of every playbook application" section explicitly codifying the W988 + W989 outcomes: (a) W988 closed correctly as not-applicable because the premise didn't match cmd_conventions.py; (b) W989 found a DIFFERENT real Pattern-2 gap than the playbook's framing assumed (W718 floor in _normalise_pr_risk_level, not a slot.update framing). The amendment codifies: "a playbook that applies mechanically without checking the premise produces fabricated work; a playbook that applies with discipline either seals the nominated gap, finds a different real gap, or stops cleanly".

Operational findings (W1001 batch — pending fix / decision)

• W984 — Autouse conftest for the bus_factor stale-kind fixture monkeypatch. W978 fix is per-test; promoting to an autouse conftest fixture would cover sibling tests that may regress when the W405 shallow-history default shifts. | tests/conftest.py or tests/test_findings_bus_factor.py | 30 min
• W985 — INFO log when W405 shallow-history drops a commit. The drop is currently silent at the git_history ingestion layer; an INFO log naming the dropped SHA + reason would help future "why is this test failing" investigations. | src/roam/index/git_stats.py | 30 min
• W986 — CLAUDE.md "first hypothesis" checklist for test-failure triage. W978 root-cause took two passes because the W405-shallow-history hypothesis wasn't first on the agent's checklist. Capture the lesson: when a test__stale_ or test__history_ test fails, the first hypothesis to check is "did W405 truncate the fixture's expected commit?". | CLAUDE.md Quality-discipline section | 30 min
• W980 / W981 — W974 UX papercuts. W980: the new AlertThreshold.level Literal[...] tightening produces a generic TypedDict error message on invalid input; a LITERAL_LEVEL_VALUES error message in _coerce_level would be more actionable. W981: the _coerce_level error message format doesn't follow LAW 4 concrete-noun-terminal vocabulary. | src/roam/commands/cmd_alerts.py | 30 min each
• W992 / W993 — W991 drift-guards. W992: add an AST lint asserting every LIKE ? in detectors.py is paired with ESCAPE '\\' — prevents the next reviewer from re-introducing an unescaped wildcard pattern. W993: an end-to-end smoke test asserting find_in_sorted_array does NOT match finXinXsortedXarray (the exact false-positive case W991 sealed). | tests/test_w992_sql_escape_drift.py + tests/test_w993_finXsortedXarray_smoke.py | 1h each
• W997 / W998 — expires ↔ expires_on field name divergence in smells_suppress YAML. W994 standardised on expires; sister suppression substrates carry expires_on or expiry. Pattern-3b parameter-name canonicalisation gap. | src/roam/policy/suppression_v2.py + sibling parsers | 1-2h

W1015-CONSOLIDATE — Pattern-2 propagation arc continuing + disclosure-hygiene class identified + YAML loader hardening memo

> CONSOLIDATE checkpoint = W1015. ~11 waves closed since W1001-CONSOLIDATE. > Headline: Pattern-2 propagation continues with W706 (cmd_ignore_findings) > + W1009 (per-finding-suppressions audit) + W1011 (cmd_alerts section-level > audit confirmation), bringing the running Pattern-2 closure count to 3 > more loader surfaces sealed this batch; a new disclosure-hygiene class > identified — W1000 sealed the strip_list_payloads warnings_out drop > via a new _ALWAYS_PRESERVED_LIST_FIELDS allow-set, defeating the previous > half-fix where partial_success=true survived but the structured warnings > array was stripped; a shared YAML loader hardening memo shipped (W1016- > RESEARCH) recommending a roll-our-own 2-phase migration (~125 LOC net > removed at 5 of 7 callsites); W996 docs the click-vocab divergence > across 7 commands as Pattern-3b parameter-name canonicalisation gap; > W1002 + W1003 fix test discipline — relative test-date offsets + > xfail-strict pin comment that survives autouse-fixture interactions; > W1015 lands the catalog _shared.py test coverage (24 tests + new > tests/test_catalog_shared.py). W886 / W890 verified already-guarded > via the W873 canonical (is_test_path None-guard is present at the > canonical site — no work needed). W494 verified clean — > test_inter_unused_return order-sensitivity audit found taint > already deterministic. Hash-stability mandate held trivially across the > batch — every fix either added new validation paths (no pre-fix envelope > bytes), preserved through-flow of an existing field (no detector output > bytes moved), or landed in test infrastructure (no source bytes).

Closed — Pattern-2 propagation continuing (W706 / W1009 / W1011 — W1015-CONSOLIDATE)

• W706 — cmd_ignore_findings Pattern-2 close. YAML-loader unknown-key path now plumbs warnings_out + flips summary.partial_success=true + surfaces an agent_contract.facts entry on unknown keys instead of silently dropping the entry. Matches the W918 envelope shape proven on cmd_alerts. Tests pass; hash-stability held (no pre-fix envelope to compare against — new validation surface). See dev/CMD-ALERTS-PATTERN-2-CASE-STUDY-2026-05-15.md for the playbook applied. | src/roam/commands/cmd_ignore_findings.py

<!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->

• W1009 — Per-finding-suppressions Pattern-2 close. Sister to W994 / W995 (smells_suppress) but on the finding-suppressions substrate. Malformed entries previously dropped silently; now partition into valid + dropped lists, surface each via indexed warnings_out entries + rollup, flip partial_success=true. Same envelope shape as W994 / W995. | src/roam/policy/finding_suppressions.py
• W1011 — cmd_alerts Pattern-2 audit confirmation. Audited cmd_alerts.py section-level loader paths post-W918 / W962 / W963 / W964 / W967 / W968 / W969; confirmed every silent-fallback surface now flows through warnings_out. No new source changes; closes the audit follow-up. Captured methodologically: the W983 playbook-propagation discipline (premise verification first) successfully reproduced the W977 close. | audit-only

Added — Disclosure-hygiene class identified (W1000 / W996 — W1015-CONSOLIDATE)

• W1000 — strip_list_payloads warnings_out preservation. Introduced _ALWAYS_PRESERVED_LIST_FIELDS allow-set in src/roam/output/formatter.py covering warnings_out (and any future Pattern-2 disclosure field that must survive the --detail-strip post-processor). Pre-fix: the partial_success=true flag survived but the structured warnings array was silently stripped when callers omitted --detail, defeating the W918 disclosure half-way. Companion new lint test asserts the allow-set covers every field touched by the Pattern-2 envelope shape. Tests pass. Surfaced during W987 playbook apply (W1001-CONSOLIDATE). | src/roam/output/formatter.py + new lint test
• W996 — Click-vocab divergence documented across 7 commands. Memo notes 7 commands where the --kind vs --type vs --metric vocabulary diverges at the CLI boundary; same Pattern-3b parameter-name canonicalisation shape as the _PARAM_ALIASES table for MCP boundary normalization. Captured as a documentation-grade audit feeding W1004 (7-cmd audit). | docs-only memo

Closed — Test discipline improvements (W1002 / W1003 / W494 — W1015-CONSOLIDATE)

• W1002 — Test-date relative offsets. Tests that hard-coded absolute dates were brittle against the autouse fixture (freeze_time) interaction; flipped to relative offsets so the test suite stays green across the year boundary without per-test override. Mirrors the W978 fixture-monkeypatch discipline. | tests/test__history_.py cluster
• W1003 — xfail-strict pin comment. Pinned the rationale for the W819-class xfail-strict flips with an inline comment noting the autouse-fixture interaction surfaced by W1002; prevents the next reader from flipping back to absolute dates. | inline comments on the xfail-strict tests
• W494 — test_inter_unused_return order-sensitivity verified clean. Audit of the taint inter-procedural unused-return analysis found the result set is deterministic across input order; no fix needed. Closes the W494 / W495 / W496 W533-bundle drive-by chain (Java leg). | audit-only

Closed — Not-applicable (W886 / W890 — W1015-CONSOLIDATE)

• W886 / W890 — is_test_path None-guard verified already-guarded at canonical. The W873-era canonical (changed_files.is_test_file) already None-guards its path argument; the W886 drive-by-2 / W890 carry-forward turn out to be redundant. No source changes; captured as audit-only closure. | audit-only (closes W886 drive-by + W890 carry-forward)

Test infrastructure (W1015 — W1015-CONSOLIDATE)

• W1015 — tests/test_catalog_shared.py (24 tests). Direct coverage of the src/roam/catalog/_shared.py canonical helpers (W864 _loc, W873 is_test_path, W877 _enclosing_symbol, W923 make_smell_finding). Pre-W1015 the helpers had only transitive coverage via per-detector tests; the new file pins the canonical contract directly so future hoists can land with a single-file regression check. 24/24 pass. | tests/test_catalog_shared.py

Research memos (W1016-RESEARCH — W1015-CONSOLIDATE)

• W1016-RESEARCH — YAML loader hardening memo. Memo evaluates the 7 YAML-loader callsites across the codebase (cmd_alerts / cmd_ignore_findings / smells_suppress / finding_suppressions / rules/loader / permit/loader / constitution/loader). Verdict: roll our own shared helper rather than depend on a third-party YAML hardening library — the surface area is small enough that the discipline rule + a thin shared parser is cheaper than a dependency. 2-phase plan: Phase 1 = W1018 (queued — extract the canonical _parse_yaml_with_warnings + _VALID_OPS-style closed-set validator); Phase 2 = W1019 (queued — migrate 5 of 7 callsites; 2 keep their bespoke parsers due to non-YAML envelope shapes). Net ~125 LOC removed across the migration. | dev/YAML-LOADER-HARDENING-2026-05-15.md <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->

Operational findings (W1015 batch — pending fix / decision)

• W1004 — 7-command click-vocab audit (W996 follow-up). Audit the --kind vs --type vs --metric boundary vocabulary across the 7 surfaced commands; recommend a canonical normalisation. | per-command audit | 2-3h
• W1005 — 3-tier vs 5-tier severity Pattern 3a divergence. Surfaced during W1011 audit — some surfaces report low/medium/high, others info/low/medium/high/critical. Same Pattern-3a shape as W596 / W631. | per-command audit | 1-2h
• W1006 — Formatter sibling preserved-fields expansion (W1000 follow-up). _ALWAYS_PRESERVED_LIST_FIELDS likely needs to expand to cover sibling fields (e.g. dropped_entries, notes); audit the disclosure-hygiene class end-to-end. | src/roam/output/formatter.py | 1-2h
• W1007 — agent_contract:[] mistake captured during W1011 audit. Some envelopes emit agent_contract:[] (empty list) instead of agent_contract:{} (empty dict); breaks consumer schema. | per-emitter sweep | 1h
• W1008 — list_counts surfacing. Pattern-2 envelopes should surface counts at the envelope root (warnings_count, dropped_count) rather than only in nested structures; helps agents check disclosure without parsing the full list. | envelope shape design + emitters | 2-3h
• W1010 — DEFERRED behind W1018. Captured as deferred pending pending the W1018 shared helper landing; revisit after Phase 1. | TBD (deferred) | TBD
• W1012 — Test-date triage. Sweep remaining hard-coded test dates for the same autouse-fixture interaction W1002 sealed. | tests/ sweep | 1-2h
• W1013 / W1014 — changed_files None-guard sweep. Sister sites to W886 / W890 — sweep the remaining helpers in changed_files for missing None-guards on path arguments. | src/roam/commands/changed_files.py | 30 min each
• W1017 — Typed wrapper plumb (W918 / W933 family follow-up). The warnings_out: list[str] | None accumulator pattern would benefit from a typed wrapper class (WarningsOut) instead of the bare-list-with-side-effects pattern. | new module + 7 callsite migrations | 2-4h
• W1018 — Phase 1 of W1016 YAML helper. Extract canonical _parse_yaml_with_warnings + closed-set validator from cmd_alerts. | src/roam/policy/_yaml_helper.py (new) | 2-3h
• W1019 — Phase 2 of W1016 YAML helper migrations (5 of 7 callsites). Blocked behind W1018. | 5 callsite migrations | 4-6h
• W1020 — Fixture-scope audit (W1002 follow-up). Audit the test fixtures for hidden autouse-scope interactions like the one W1002 sealed. | tests/conftest.py + per-test scope audit | 2-3h
• W1021 — camel_split location verify (W901 memo drift). Memo claim about _camel_split canonical location may differ from current source; W929 already moved the canonical to tfidf.py but W901's __all__ export note may be stale. Verify and refresh. | docs / memo update | 30 min
• W1022 / W1023 / W1024 — Small _shared.py polish (W1015 drive-bys). Each is a sub-30-min polish item surfaced while writing the test file: better type annotations, docstring tightening, internal helper consolidation. | src/roam/catalog/_shared.py | 30 min each
• W1025 — Alerts section-level Pattern-2 sibling (W1011 audit follow-up). One sister loader path in cmd_alerts still emits the section-level envelope without the new warnings_out plumbing; trivial extension once W1018 lands. | src/roam/commands/cmd_alerts.py | 30 min

• W1000 — REAL: strip_list_payloads drops warnings_out without --detail. Surfaced during the W987 playbook apply — the envelope post-processor in roam.output.formatter strips warnings_out from the response without --detail flag, defeating the Pattern-2 disclosure. Sealed by W1000 (W1015-CONSOLIDATE) — see "Disclosure-hygiene class identified" section above. | src/roam/output/formatter.py | shipped

W1042-CONSOLIDATE — Shared YAML helper arc lands + Pattern-2 propagation continues + cargo-cult or "" family substantially clean

> CONSOLIDATE checkpoint = W1042. ~18 waves closed since W1015-CONSOLIDATE. > Headline: the W1016-RESEARCH shared YAML helper landed (W1018) and > migrated through 4 of 5 planned Phase 2 callsites (W1019a/c/d/e; W1019b > in flight), each migration revealing a real helper-contract gap that was > sealed as a discrete follow-up wave — W1035 (JSON parse-error wording), > W1040 (PyYAML strict-timestamp force_tiny_parser kwarg), W1031 (typed > overload), W1043 (WarningsOut type alias). This is the BAIL-AND-SEAL > discipline working as designed — the migration agent bailed cleanly the > moment the helper's contract was insufficient, the gap was sealed as > a separate W, then the migration re-dispatched clean. Pattern-2 > propagation continued with 4 more loader surfaces Pattern-2-fied: > W1017 (load_per_finding_suppressions_typed warnings_out plumb), > W1025 (cmd_alerts thresholds-section sibling), W1032 > (load_suppressions + load_suppressions_typed deeper close + helper > migration), W1042 (sarif._load_suppressions_typed warnings_out > plumb). Running Pattern-2 loader-site total is now ~33-34 sites > sealed end-to-end; W1039-RESEARCH evaluated whether to evolve the > envelope shape for Python 3.13+ and concluded STAY — the current > playbook is the right one. Cargo-cult or "" cleanup swept 11 > sites across the W1029 batch (10 sites + 3 helpers None-guarded) + > W1034 (1 more in causal_graph.py); combined with W1013/W1014 the > family stands at 14 cargo-cult removals across 9 files and W907 > false-hedge anti-pattern enforcement is now mostly clean. **catalog/* > __all__ discipline landed via W1033 (_shared.py) + W1037 (5 > sibling catalog modules) for a total of 6 catalog modules with > explicit __all__ declarations, deterring the same cross-module > private-name-import pattern W901 originally captured. W1026 back-fill > retroactively annotated the W1016-RESEARCH memo with the W1018 tiebreaks > observed during Phase 1 implementation. Hash-stability mandate held > trivially across the batch** — every fix either added new validation > paths (no pre-fix envelope bytes), tightened type surfaces (no runtime > behavior delta), or hoisted helper internals (no detector output bytes > moved).

Added — Shared YAML helper arc (W1018 / W1019a/c/d/e / W1031 / W1035 / W1040 / W1043 — W1042-CONSOLIDATE)

• W1018 — load_yaml_with_warnings helper shipped. New module + 23 tests. Phase 1 of the W1016-RESEARCH 2-phase plan. Canonical _parse_yaml_with_warnings extracted from cmd_alerts; closed-set validator pattern generalised. Initially unused — landed alongside Phase 2 migration to keep the abstraction honest (per CLAUDE.md "Don't TypedDict a boundary you don't validate" discipline, the helper's shape was iterated against real migrations rather than designed in isolation).
• W1019a / W1019c / W1019d / W1019e — Phase 2 migrations. 4 of 5 planned callsites migrated to the new helper (cmd_ignore_findings, smells_suppress, per_finding_suppressions, rules/loader); W1019b in flight (partial-completion noted — re-dispatch after the W1040 extension landed). The migration approach — bail cleanly on first contract gap, capture the gap as a discrete wave, re-dispatch after seal — proved out on three separate occasions (W1035 / W1040 / W1031).
• W1035 — parse_error_label kwarg. Helper contract extension surfaced during Phase 2: the helper's JSON-parse-error wording assumed YAML context, breaking the JSON envelope shape for callers that needed a different label. Sealed as a discrete kwarg + 1 test.
• W1040 — force_tiny_parser kwarg. Helper contract extension surfaced during the smells_suppress migration: PyYAML's strict-timestamp parsing was inverting W994's EXPIRES_FMT discipline (yaml.safe_load accepts unbounded ISO formats; the tiny fallback parser is the EXPIRES_FMT-compatible path). Fix: kwarg lets callers force the tiny parser even when PyYAML is installed. Unblocks the W994 / W995 envelope shape preservation across the smells_suppress migration.
• W1031 — Typed overload for load_yaml_with_warnings. Phase 2 surface: warnings_out typed return signature variant for callers consuming the helper through a TypedDict-fronted loader. Companion to the W933 / W919 / W966 typing discipline.
• W1043 — WarningsOut type alias. Concurrent with this wave; finalises the W1017 typed wrapper plumb as a TypeAlias = list[str] declaration at the canonical boundary (no behavior change; readability + LSP hint quality only).

Pattern-2 propagation (W1017 / W1025 / W1032 / W1042 — W1042-CONSOLIDATE)

• W1017 — load_per_finding_suppressions_typed warnings_out plumb. Per-finding-suppressions typed loader sister to W1009; full Pattern-2 envelope shape (warnings_out + partial_success=true + agent_contract.facts) now flows through the typed surface. Matches the W918 / W994 / W995 / W1009 shape.
• W1025 — cmd_alerts thresholds-section Pattern-2 sibling closed. The one sister loader path captured during the W1011 audit (W1015-CONSOLIDATE pending) — the thresholds-section envelope now plumbs warnings_out alongside the section-level surfaces W918 / W962 / W963 / W964 / W967 / W968 / W969 had sealed. Trivial extension once W1018 landed; final cmd_alerts.py Pattern-2 sibling sealed.
• W1032 — load_suppressions + load_suppressions_typed deeper Pattern-2 + helper migration. Both loaders now flow through load_yaml_with_warnings; the deeper-Pattern-2 close covers the previously-untreated "valid YAML, semantically-incoherent payload" surface (sister to W1009's malformed-entry partition).
• W1042 — sarif._load_suppressions_typed warnings_out plumb. Sarif emitter's private typed loader migrated to the canonical shape; closes the last loader-site sibling that was still emitting the typed envelope without warnings_out flow. Running Pattern-2 loader-site total ~33-34 sealed across the W918 → W1042 arc.

Changed — Cargo-cult or "" cleanup (W1029 / W1034 — W1042-CONSOLIDATE)

• W1029 — 10 sites cleaned + 3 helpers None-guarded. Audit pass found 10 sites doing path or "" / value or "" defensive-coercion in spots where the upstream contract already guaranteed non-None (cargo-culted from genuinely-defensive sister sites). 3 helpers were tightened with explicit None-guards instead (changed_files-family + sibling utility modules). Same shape as the W907 false-hedge anti-pattern audit; preserves the rule "verify the precondition before defending against it".
• W1034 — causal_graph.py:713 cargo-cult cleanup. Drive-by from W1029 — one more or "" site at the causal-graph layer that the W1029 sweep missed. Combined W1013/W1014 + W1029 + W1034 = 14 cargo-cult removals across 9 files. W907 false-hedge enforcement is now mostly clean across the codebase.

Added — catalog/* __all__ discipline (W1033 / W1037 — W1042-CONSOLIDATE)

• W1033 — src/roam/catalog/_shared.py __all__ declaration. Explicit __all__ listing the canonical helpers (_loc, is_test_path, _enclosing_symbol, make_smell_finding, etc.) per the W901 cross-module private-name-import discipline. Closes a sub-30-min polish item from the W1015 batch.
• W1037 — 5 sibling catalog modules gain __all__. Uniform discipline applied across smells.py + detectors.py + parallel_hierarchy.py + clones_cross_layer.py + type_switch.py. 6 catalog modules total with explicit __all__ declarations (W1033 + W1037); deters the W901-class private-name-import pattern across the catalog surface.

Research memos (W1039-RESEARCH — W1042-CONSOLIDATE)

• W1039-RESEARCH — Python 3.13+ Pattern-2 evolution memo: STAY verdict. Evaluated whether to evolve the W918 envelope shape ahead of Python 3.13's expanded type-narrowing surface (PEP 695 / PEP 705 / PEP 692 interactions with the warnings_out: list[str] | None accumulator pattern). Verdict: STAY the course. The current playbook (bare-list accumulator + partial_success flag + agent_contract.facts disclosure) is structurally cheaper than a TypedDict-fronted alternative, AND the W933 / W966 discipline rule explicitly counsels against TypedDict-fronting at this boundary because the accumulator is populated via list.append(arbitrary) rather than literal construction. The W1043 WarningsOut type alias is the maximum-tightening that stays inside the discipline.

Memo back-fills (W1026 — W1042-CONSOLIDATE)

• W1026 — W1016-RESEARCH memo back-fill with W1018 tiebreaks. The original W1016 memo captured the decision verdict (roll our own, 2-phase plan) but did not capture the tiebreaks the agent actually used when implementing W1018 (e.g. why the helper exposes a parse_error_label kwarg ahead of any caller needing it — captured in W1035). Back-fill folds the W1018 / W1035 / W1040 / W1031 / W1043 evolution into the memo's "Phase 1 implementation notes" section so the next reader sees the iteration, not just the verdict.

Operational findings (W1042 batch — pending fix / decision)

• W1036 — 4 sibling _parse_simple_yaml loaders still bespoke (W1018 / W1019 follow-up). 4 loader sites carry their own tiny YAML parsers (smells_suppress's original, finding_suppress's original, two more in policy/ and rules/). Each is a candidate for migration to load_yaml_with_warnings once the helper has a force_tiny_parser path proven through one more callsite. | per-loader audit + migration | 2-3h
• W1038 — _extract_typed helper hoist (W1031 follow-up). The typed-overload path carries a _extract_typed adapter at each callsite; once 2+ callsites stabilise, the adapter should hoist to a shared utility. | new helper or _yaml_helper.py extension | 1h
• W1041 — clones_cross_layer.py __all__ divergence (W1037 follow-up). The W1037 sweep used a slightly different __all__ ordering convention than the W1033 baseline (alphabetical vs declaration-order); resolve. | src/roam/catalog/clones_cross_layer.py | 15 min
• W1042 — sarif._load_suppressions_typed typed-loader migration (concurrent with this wave; if not yet shipped, queued). | src/roam/output/sarif.py | 30 min
• W1043 — WarningsOut type alias (concurrent with this wave; if not yet shipped, queued). | src/roam/_types.py (new) or canonical boundary module | 30 min
• W1044 — DEFERRED behind W1019b re-dispatch. Captured as deferred pending the W1019b re-dispatch landing post-W1040. | TBD (deferred) | TBD
• W1004 — 7-cmd click-vocab audit (W996 follow-up, carry-from-W1015). Still queued. | per-command audit | 2-3h
• W1005 — 3-tier vs 5-tier severity Pattern 3a divergence (W1011 audit follow-up, carry-from-W1015). Still queued. | per-command audit | 1-2h

---

> TWO CRITICAL edge-attribution fixes seal the call/import edge family end-to-end (W708 + W742) + suppression family migration extended through Phase C-1 (W722 → W723 → W736 → W737 → W738) + Pattern-3a canonical-rank consolidation third axis CLOSED (W596 confidence + W631 risk + W640 alerts level_order fold + W648 zero-slipped audit + W649 alerts lowercase) + bare-except discipline structurally CLOSED (allowlist 9 → 3 via W660 / W665 / W677 / W678 / W679 / W740 + W662 / W746 _GUARDED_DIRS 4 → 12 + W707 dead-code REAL BUG seal) + wheel-bundling thread CLOSED (W664 LIVE BUG caught + W668 as_file audit + W624 importlib.resources migration + W642 / W643 fallback removals) + MCP wrapper P0 batch (W670 alias regression + W671 cold-start guard + W672 surface sync + W606 collision lint + W607 / W636 wrapper refactor + W695 --card smoke) + smell catalog reached 20 detectors (W601-W605 / W639 / W646 / W647 / W650 / W705 / W720) + hygiene drive-bys (W682 / W683 / W685 / W689 / W690 / W697 / W699 / W702) (~20-completion batch behind W755-CONSOLIDATE, 2026-05-15). > The headline is TWO systemic edge-attribution correctness fixes that together seal the family. W708 (call-edge mis-attribution): _store_symbols + _merge_existing_symbols in src/roam/index/indexer.py:551,1192 omitted line_end from all_symbol_rows. The resolver's le > 0 guard always failed → syms[0] fallback for every method ref. Repo-wide 95% mis-attribution reduction (2715 → 147 silently corrupted edges). W742 (phantom import-edge mis-attribution): _closest_symbol in src/roam/index/relations.py:488-496,848-898 was returning syms[0] for kind='import' refs that couldn't be resolved to an enclosing symbol — manufacturing 18 phantom import edges on _format_count alone plus 6 transitive side-effects. Fix: optional kind parameter on _closest_symbol; returns None for kind=='import' instead of falling through. New invariant test in tests/test_relations.py:167-225. Together, every detector that reads edges (taint / side_effects / critique / dead / smells / vibe-check / ai-rot) now consumes correct edges. Suppression family migrated through Phase C-1 (W691 schema unification → W692 dataclass src/roam/policy/suppression_v2.py 312-line _SuppressionBase + 3 variants → W722 Phase B-a smells typed companion → W723 Phase B-b finding_suppress + sarif → W736 Phase C-1a sarif _load_suppressions migrated → W737 Phase C-1b cmd_smells.load_smells_suppressions migrated → W738 Phase C-1c BAILED on cmd_triage for three malformed-input divergences but MIGRATED suppression.save_suppression internal dedup; new tests/test_w738_suppression_wire_format.py 8/8 pass). Pattern-3a third-axis close: W596 src/roam/output/confidence.py (15 sites) + W631 src/roam/output/risk.py (4-tier critical/high/medium/low + moderate→medium alias) + W640 cmd_alerts._LEVEL_ORDER folded into severity_rank() + W648 AST audit ZERO slipped tables + W649 cmd_alerts UPPER → lowercase per W547. Bare-except discipline shipped end-to-end: W660 _find_workspace_root narrowed; W661 catalog/detectors.py production loop fail-loud; W662 AST drift-guard; W665 / W677 / W678 / W679 / W740 narrowed individual sites (allowlist 9 → 3); W707 found + sealed a REAL BUG (_serialize_suppressions dead-code on the first flag); W746 extended _GUARDED_DIRS 4 → 12 to cover substrate modules. Wheel-bundling thread closed: W664 __init__.py package-data drift-guard CAUGHT A LIVE W643-class bug on first run (roam.languages.extractors was missing its __init__.py); W668 as_file() audit; W624 migrated mcp_server.py:14569 mcp --card handler to importlib.resources; W642 / W643 removed dead triple-parent fallback. MCP wrapper P0 batch: W670 P0.1 fixed roam_plan file_path alias regression by moving _wrap_with_alias_normalization BEFORE the preset filter; W671 P0.2 added _INLINE_RESPONSE_TOOLS frozenset cold-start exemption for roam_catalog; W672 P0.3 synced scripts/sync_surface_counts --write to live 238/231/224; W606 added AST lint for canonical-positional collision; W607 decomposed _wrap_with_alias_normalization into 3 helpers; W636 collapsed _sync vs _async wrapper closure duplication; W695 added --card CLI smoke test. Smell catalog reached 20 detectors: W601 switch-statement (7 findings; surfaced REAL refactor candidate _create_extractor 23-arm switch); W602 temporal-coupling (10 findings); W603 magic-numbers (495 findings); W604 boolean-parameter (0 findings); W605 comment-density TODO/FIXME/XXX/HACK; W639 cross-detector empty-corpus smoke; W646 W699 DOGFOOD refactor _format_count (cluster finding led to W708 + W742); W647 symbol-centric temporal-coupling rollup (surfaced W708 false positive); W650 block-comment TODO/FIXME C/Java/JS; W705 unified _CommentSyntax (21 languages); W720 comment-density extended to hcl + apex. Hygiene drive-bys: W682 README CLI table evidence-oscal row; W683 .gitattributes 13 → 49 lines (* text=auto eol=lf + 26 binary extensions); W685 README CLI table header auto-count assertion; W689 .editorconfig mirroring .gitattributes; W690 dev-doc note for pytest on Windows; W697 extras-gate on README CLI command-count check; W699 DOGFOOD refactor _format_count cluster finding; W702 _DEPRECATED_COMMANDS AST-literal contract test. Hash-stability mandate held 31/31 byte-identical across every source wave.

Fixed — CRITICAL edge-attribution family CLOSED (W708 call-edge mis-attribution + W742 phantom import-edge mis-attribution)

• W708 — CRITICAL: Python call-edge mis-attribution fixed. _store_symbols (src/roam/index/indexer.py:551) + _merge_existing_symbols (src/roam/index/indexer.py:1192) omitted line_end from all_symbol_rows. Resolver's le > 0 guard always failed → syms[0] fallback for every method ref. Repo-wide 95% mis-attribution reduction (2715 → 147 corrupted edges). Every detector that reads edges (taint / side_effects / critique / dead / smells / vibe-check / ai-rot) silently consumed corrupted edges pre-fix. Surfaced by W647's symbol-centric temporal-coupling rollup (a clustered false positive pointed straight at the mis-attribution).
• W742 — CRITICAL: Phantom import-edge mis-attribution to first symbol fixed. _closest_symbol in src/roam/index/relations.py:488-496,848-898 was returning syms[0] for kind='import' refs that did not resolve to an enclosing symbol. Result: 18 phantom import edges on _format_count + 6 transitive side-effects. Fix: optional kind parameter on _closest_symbol returns None for kind=='import' instead of falling through. New invariant test in tests/test_relations.py:167-225 pins the contract. Pairs with W708 to close the edge-attribution family end-to-end across BOTH call-edge AND import-edge resolution.
• W707 — REAL BUG: _serialize_suppressions dead-code on first flag. Investigated the suppression serializer; found a dead first flag that was set but never read. Removed; regression test pins the call-site count at zero. Pairs with W691 schema unification + W722 / W723 phased migration.
• W740 — _load_project_config bare-except narrowed. Continues the W662 drift-guard scoping; allowlist 4 → 3.

Fixed — Suppression family migrated through Phase C-1 (W722 / W723 / W736 / W737 / W738)

• W722 Phase B-a — Smell-suppression typed companion. load_smells_suppressions_typed() returns KindSymbolSuppression records bridging the legacy smell-suppression substrate (W658) into the W692 dataclass surface.
• W723 Phase B-b — finding_suppress + sarif typed. Extended Phase B to cmd_finding_suppress + the sarif emitter; both consume FindingIdSuppression records.
• W736 Phase C-1a — sarif _load_suppressions migrated. First Phase C migration — the sarif emitter's private loader now flows through the canonical src/roam/policy/suppression_v2.py parser.
• W737 Phase C-1b — cmd_smells.load_smells_suppressions migrated. Second Phase C migration — cmd_smells consumes the canonical loader.
• W738 Phase C-1c — cmd_triage BAILED, suppression.save_suppression MIGRATED. Investigated cmd_triage for Phase C-1c migration; found three malformed-input divergences between the legacy parser's lenient behavior and the canonical loader's closed-schema validation — bailed on cmd_triage migration to keep the legacy behavior on the user-facing triage surface, but migrated the internal suppression.save_suppression dedup path. New tests/test_w738_suppression_wire_format.py pins the wire format (8/8 pass).

Changed — Pattern-3a canonical-rank consolidation third axis CLOSED + alerts surface canonicalised (W596 / W631 / W640 / W648 / W649)

• W596 — Confidence-level rank canonical helper. New src/roam/output/confidence.py. 15 sites migrated. Pairs with W564 severity-rank (10 sites) to give Pattern-3a its second canonical axis end-to-end.
• W631 — Risk rank canonical helper. New src/roam/output/risk.py. 4-tier closed enum (critical / high / medium / low) with moderate → medium alias for backward compat. Third Pattern-3a axis CLOSED — combined with W596 confidence + W564 severity, every rank surface in roam now flows through canonical helpers with AST drift-guards.
• W640 — cmd_alerts._LEVEL_ORDER folded into severity_rank(). Sort key now uses -severity_rank(lowercase) instead of a private rank table. Drift-guard regex broadened so the canonicalisation is enforced for any future _LEVEL_ORDER clones.
• W648 — AST audit confirmed ZERO slipped rank tables. Audited the entire src/ tree for inline rank tables that would bypass the canonical helpers. Zero slipped. Pattern-3a structurally closed for real, not just-in-name.
• W649 — cmd_alerts UPPER → lowercase canonicalisation. Per W547 closed-severity-vocab contract (CRITICAL / WARNING / INFO → lowercase). Pairs with W640 to give the alerts surface its full canonical-vocab discipline.

Changed — Bare-except discipline structurally CLOSED (W660 / W662 / W665 / W677 / W678 / W679 / W746)

• W660 — _find_workspace_root bare-except narrowed. Same site as the _load_project_config narrowing; specific-exception classification only.
• W661 — catalog/detectors.py production loop fail-loud. Classifies NameError / ImportError / AttributeError / TypeError as RuntimeError (re-raise) vs sqlite3.Error (swallow + log). Pair with W653 source fix + W662 drift-guard for end-to-end discipline.
• W662 — AST drift-guard banning bare except Exception: continue/pass. Initial _GUARDED_DIRS was 4 directories.
• W665 — 3 specific bare-except sites narrowed. Continues the W662 scoping.
• W677 — formatter.py:420,905 narrowed. Continues W665 scoping.
• W678 — taint_engine.py:133 narrowed once the parser stabilised. Was previously held back by the parser flux; W662 stabilisation unblocked the narrowing.
• W679 — detectors.py:4165 narrowed. Closed-set sqlite3.Error / KeyError / TypeError; allowlist 3 → 2.
• W746 — Extended W662 _GUARDED_DIRS 4 → 12 to substrate modules. The drift-guard now covers the agent-OS substrate packages (constitution / modes / runs / leases / memory / pr-bundles / laws / agents_md) in addition to the original 4. Allowlist now structurally pinned at 3 sites total with the substrate covered.

Changed — Wheel-bundling thread CLOSED (W624 / W642 / W643 / W664 / W668)

• W624 — Migrated mcp_server.py:14569 mcp --card handler to importlib.resources. Continues the W554 / W535 / W610 discipline thread; parents[N] path-juggling replaced with importlib.resources.files("roam") / "mcp-server-card.json".
• W642 — Removed triple-parent fallback from mcp --card handler. -19 LOC. W624 already migrated the resolution; the parents[3] fallback was dead code after that.
• W643 — Grepped remaining Path(__file__).parent resource loads. Audit pass confirming the importlib.resources migration is complete across src/.
• W664 — __init__.py package-data drift-guard. CAUGHT A LIVE W643-class bug on first run: roam.languages.extractors was missing its __init__.py, meaning the package was not actually shipped in the wheel despite being referenced via package-data.
• W668 — Audited as_file() callers for path-captured-outside-with anti-pattern. 4 sites fixed; drift-guard pins the pattern at lint time.

Changed — MCP wrapper P0 batch + refactor chain (W606 / W607 / W636 / W670 / W671 / W672 / W695)

• W606 — AST lint for canonical-positional collision. 4 new tests catching the pre-W595 crash class at PR time.
• W607 — Refactored _wrap_with_alias_normalization into 3 helpers. _collect_alias_candidates + _build_merged_signature + _build_merged_annotations. 130 → 50 lines + 7 unit tests.
• W636 — Collapsed _sync vs _async wrapper closure duplication. New shared _prepare_kwargs helper + branched closure pattern. 33 → 28 lines and duplicate-body anti-pattern eliminated.
• W670 P0.1 — roam_plan file_path alias regression fixed. Moved _wrap_with_alias_normalization BEFORE the preset filter so the alias normalisation layer wraps every tool, not just those that survive the filter. User-flagged P0.
• W671 P0.2 — roam_catalog cold-start auto-handle exemption. New _INLINE_RESPONSE_TOOLS frozenset exempts roam_catalog from _wrap_with_handle_off — cold-start call returns inline. User-flagged P0.
• W672 P0.3 — scripts/sync_surface_counts --write synced to live 238/231/224. README + CLAUDE.md + llms-install.md + both mcp-server-card.json copies + server.json + landing-page HTML all carry the live counts. User-flagged P0.
• W695 — --card CLI smoke test (2 tests). Pins the mcp --card handler against silent regression — W624 + W642 migrations now have an end-to-end CLI test.

Added — Smell catalog reached 20 detectors (W601-W605 / W639 / W646 / W647 / W650 / W699 / W705 / W720)

• W601 — switch-statement smell detector. 7 findings on roam-code; surfaced REAL refactor candidate _create_extractor 23-arm switch (sealed by W646).
• W602 — temporal-coupling smell detector. 10 findings on roam-code. Continues W370c.
• W603 — magic-numbers smell detector. 495 findings on roam-code.
• W604 — boolean-parameter smell detector. 0 findings on roam-code — the discipline already holds.
• W605 — comment-density smell detector (TODO / FIXME / XXX / HACK).
• W639 — Cross-detector empty-corpus smoke test. Guards 54 detectors against silent import errors after concurrent merges.
• W646 — REFACTOR: _create_extractor 105 → 17 lines via _LANGUAGE_EXTRACTORS dispatch dict. Eat-our-own-dogfood seal of the W601 finding.
• W647 — Symbol-centric temporal-coupling rollup. 10 pair findings → 5 cluster findings on roam-code; cmd_health.health clustered. Surfaced the false positive that drove the W708 critical fix.
• W650 — Block-comment TODO/FIXME detection extended to C/Java/JS. Coverage now includes C-family + CSS block comments alongside the existing line-comment detection.
• W699 — DOGFOOD: Refactor _format_count. Cluster finding that surfaced the W708 + W742 mis-attribution fixes downstream.
• W705 — Unified _CommentSyntax record. Comment-density smell coverage 14 → 21 languages.
• W720 — Comment-density extended to hcl + apex. Continues the W705 unification.

Added — Hygiene drive-bys (W682 / W683 / W685 / W689 / W690 / W697 / W702)

• W682 — README CLI table now includes the evidence-oscal row. Closes the W672 gap audit.
• **W683 — .gitattributes extended (* text=auto eol=lf + 26 binary extensions).** Closes the long-tail CRLF-on-Windows wheel-smoke failure class.
• W685 — README CLI table header-count assertion via test_readme_cli_command_count_matches_source. Fails the build if the README header drifts from the live source count.
• W689 — .editorconfig mirroring .gitattributes. Editors get a single source of truth on line-endings + charset.
• W690 — Dev-doc note for pytest on Windows. Captures the pytest-xdist + Windows quirk surfaced during the W708 + W742 validation pass.
• W697 — Extra-gate to test_readme_covers_all_canonical_cli_commands. Auto-allowlist from cli._DEPRECATED_COMMANDS so newly-deprecated commands flow through automatically.
• W702 — _DEPRECATED_COMMANDS AST-literal contract test. Pins the deprecation-list shape — any drift fails at lint time.

Changed — W755 consolidation pass

• W755 — CHANGELOG / HANDOVER / BACKLOG / SESSION-SNAPSHOT refresh for the ~20-completion batch behind W733. Docs-only; hash-stability mandate held trivially.

> Seventeen-completion batch folded in behind the W698-CONSOLIDATE consolidation. The headline is W708's critical silent-bug seal: Python call-edge mis-attribution was silently corrupting every detector that reads edges (taint, side_effects, critique, dead, smells, vibe-check, ai-rot). Root cause was indexer.py:551 + indexer.py:1192 omitting line_end from all_symbol_rows, which collapsed per-call resolution to per-symbol. Post-fix: _format_count non-import edges drop 78 → 0 on roam-code; repo-wide 2715 → 147 (95% reduction). Validation in flight (W709). The W647 symbol-centric temporal-coupling rollup (10 pair → 5 cluster findings; cmd_health.health clustered) is what surfaced the false positive driving the W708 fix. Suppression family phased close: W691 unified .roam/suppressions.json schema between finding_suppress + sarif readers (closing the W676-found latent bug); W692 Phase A shipped the discriminated-union dataclass at src/roam/policy/suppression_v2.py; W722 Phase B-a added the load_smells_suppressions_typed() companion (KindSymbolSuppression internal); W693 added cross-loader compat across 5 suppression substrates. W723 Phase B-b in flight; W724 Phase C queued. Comment-density smell expansion: W705 unified _CommentSyntax record taking coverage 14 → 21 languages; W720 extended to hcl + apex; W650 extended detection to / / block comments (C-family + CSS). Hygiene wave: W689 added .editorconfig (23 lines) mirroring .gitattributes EOL/charset/binary rules; W685 pinned README CLI table header to (all 231) with auto-count + test_readme_cli_command_count_matches_source; W695 added --card CLI smoke (2 tests); W697 added README CLI test extras-gate (auto-allowlist from cli._DEPRECATED_COMMANDS); W702 added _DEPRECATED_COMMANDS AST-literal contract test. Small cleanups: W642 removed triple-parent fallback from mcp --card handler (-19 LOC); W649 canonicalised cmd_alerts UPPER → lower per W547 contract; W707 removed _serialize_suppressions dead code + regression test. Hash-stability 31/31 byte-identical held across every source wave.

Fixed — CRITICAL silent call-edge mis-attribution (W708) + suppression schema unified (W691) + cmd_alerts canonical lowercase (W649)

• W708 — CRITICAL: Python call-edge mis-attribution fixed. indexer.py:551 + indexer.py:1192 omitted line_end from all_symbol_rows, collapsing per-call edge resolution to per-symbol. Result: every detector reading edges (taint, side_effects, critique, dead, smells, vibe-check, ai-rot) silently consumed corrupted edges. Post-fix: _format_count non-import edges drop 78 → 0 on roam-code itself; repo-wide 2715 → 147 (95% reduction). Surfaced by W647's symbol-centric temporal-coupling rollup (a clustered false positive pointed straight at the mis-attribution). Validation in flight (W709) — the impact crosses every detector family so the validation pass is broad.
• W691 — .roam/suppressions.json schema unified between finding_suppress + sarif readers. Closes the W676-found CRITICAL latent bug (two readers consuming the same file with incompatible shape contracts — suppressions silently applied to one detector and not the other). First phase of the W691 → W692 → W722 phased-close family.
• W649 — cmd_alerts UPPER-case → lowercase canonicalisation. Brings the alerts surface in line with the W547 closed-severity-vocab contract. Pairs with W640 (cmd_alerts _LEVEL_ORDER fold) to give the alerts surface its full canonical-vocab discipline.
• W642 — Removed triple-parent fallback from mcp --card handler. -19 LOC. Continuation of the W624 importlib.resources migration; the parents[3] fallback was dead code after W624.

Added — Suppression family phased close + symbol-centric temporal-coupling rollup + comment-density 14→21 languages + hygiene wave

• W692 Phase A — Suppression discriminated-union dataclass. Shipped at src/roam/policy/suppression_v2.py. Closed-schema typed surface that replaces the prior shape-divergent dict-based parsers. Pairs with W691 (schema unification) + W722 Phase B-a (typed loader).
• W722 Phase B-a — load_smells_suppressions_typed() companion. KindSymbolSuppression internal type bridging the legacy smell-suppression substrate (W658) into the W692 dataclass surface. W723 Phase B-b in flight; W724 Phase C queued.
• W693 — Cross-loader compat test for 5 suppression substrates. Pins shape parity across the family — any future schema drift fails the compat test before it ships.
• W647 — Symbol-centric temporal-coupling rollup. Replaces the prior pair-by-pair rollup with a symbol-clustered view. 10 pair findings → 5 cluster findings on roam-code; cmd_health.health clustered. This is the rollup that surfaced the W708 critical call-edge mis-attribution — the apparent "temporal coupling" between two symbols turned out to be a mis-attributed call edge.
• W705 — Unified _CommentSyntax record. Comment-density smell coverage 14 → 21 languages. Closed-schema language record replaces the prior per-language ad-hoc syntax constants.
• W720 — Comment-density extended to hcl + apex. Continues the W705 unification.
• **W650 — Comment-density extended to / / block comments.** Coverage now includes C-family + CSS block comments alongside the existing line-comment detection.
• W689 — .editorconfig added (23 lines). Mirrors .gitattributes EOL/charset/binary rules; pairs with the W683 .gitattributes extension to give Windows/Linux/Mac editors a single source of truth on line-endings + charset.
• W685 — README CLI table header pin "(all 231)" with auto-count + test_readme_cli_command_count_matches_source. Smart 231-canonical choice (matches the canonical count, not the raw 238 command count). The new test fails the build if the README header drifts from the live source count.
• W695 — --card CLI smoke test (2 tests). Pins the mcp --card handler against silent regression — the W624 + W642 migrations now have an end-to-end CLI test.
• W697 — README CLI test extras-gate. Auto-allowlist from cli._DEPRECATED_COMMANDS so newly-deprecated commands flow through the gate automatically.
• W702 — _DEPRECATED_COMMANDS AST-literal contract test. Pins the deprecation-list shape — any drift fails at lint time.

Changed — Small cleanups (W707 dead-code + W698 consolidation)

• W707 — _serialize_suppressions dead-code cleanup + regression test. The serializer was unreachable after the W691 schema unification; the regression test pins the call-site count at zero.
• W733 — CHANGELOG/HANDOVER/BACKLOG/SESSION-SNAPSHOT refresh for W691-W722 batch (this consolidation). Docs-only; hash-stability mandate held trivially.

> P0 user-flagged regression batch fixed (W670/W671/W672/W682) + bare-except discipline shipped end-to-end (W653 real bug + W661/W662 fail-loud guards + W665/W677 narrowing, allowlist 9→4) + wheel-bundling discipline COMPLETE (W664 LIVE BUG caught + W668 as_file audit + W642 triple-parent removed) + smell-suppression substrate (W658) + CRITICAL latent bug surfaced in .roam/suppressions.json (W676 → W691 in flight) + W646 eat-our-own-dogfood (W601 cleared) + W683/W685 hygiene (W670 / W671 / W672 / W682 / W683 / W685 / W642 / W646 / W653 / W661 / W662 / W664 / W665 / W668 / W658 / W676 / W677 batch, 2026-05-15). > Sixteen-completion batch folded in behind the W657-CONSOLIDATE consolidation. The headline is the P0 batch closure (4 user-flagged regressions sealed): W670 P0.1 moved _wrap_with_alias_normalization before the preset filter so roam_plan no longer drops the file_path alias on filtered presets; W671 P0.2 added a _INLINE_RESPONSE_TOOLS frozenset that exempts roam_catalog from the auto-handle wrapper so the cold-start catalog call returns inline instead of through a never-completed handle; W672 P0.3 synced 8 files to the live 238 commands · 231 canonical · 224 mcp tools counts (the auto-derive path via dev/build_readme_counts.py --apply); W682 P0.3-followup added the evidence-oscal row to the README CLI table. Bare-except discipline shipped end-to-end: W653 fixed a REAL bug in run_all_detectors — bare-except was swallowing NameError/ImportError/AttributeError/TypeError classifying-bugs as if they were per-detector failures; now they propagate as RuntimeError and only sqlite3.Error is swallowed+logged; W662 added an AST drift-guard banning bare-except in detector modules (9 sites grandfathered + 10/10 tests pass); W661 applied the fail-loud discipline to the catalog/detectors production loop (8 new tests); W665 narrowed 3 bare-except sites (allowlist 9→6); W677 narrowed 2 more (allowlist 6→4). Wheel-bundling discipline COMPLETE: W664 added a __init__.py package-data drift-guard that CAUGHT A LIVE W643-class bug on first run (roam.languages.extractors had a missing __init__.py); W668 audited as_file() callers + sealed the pattern with 4 fixes + a drift-guard; W642 removed the triple-parent fallback from the mcp --card handler (-19 LOC; W624 already migrated the resolution to importlib.resources so the fallback was dead code). Smell-suppression substrate landed (W658): 225-line module + 17 tests for .roam/smells.suppress.yml. CRITICAL latent bug surfaced (W676): suppression-parser audit found 4 parsers (not 3) with incompatible schemas in .roam/suppressions.json — two readers consume the same file with different shapes, so suppressions silently apply to one detector and not the other (W691 in flight to seal). W646 eat-our-own-dogfood: refactored _create_extractor from 105 → 17 lines via a _LANGUAGE_EXTRACTORS dispatch dict — cleared roam's own W601 finding on itself (first time the smell catalog caught a true positive on roam-code AND the refactor sealed it within the same week). W683 / W685 hygiene: .gitattributes extended 13 → 49 lines (eol=lf + 26 binary rules); README CLI table header pinned to "(all 231)" matching the 231-canonical count. Hash-stability 31/31 byte-identical held across every source wave.

Added — Smell-suppression substrate + bare-except AST drift-guard + __init__.py wheel drift-guard (W658/W662/W664 batch)

• W658 — .roam/smells.suppress.yml smell-suppression substrate. 225-line module + 17 tests. First-class suppression surface for the smell catalog; suppressions are scoped per detector + per path glob with a deterministic match order. Pairs with the W370c 5-smell expansion (W601-W605 from the prior batch) to give the catalog its operator-facing escape hatch.
• W662 — AST drift-guard banning bare-except in detector modules. 9 sites grandfathered into a _PRE_W662_PENDING allowlist (5 of those subsequently narrowed in W665 + W677, allowlist now 4). 10/10 tests pass. Closes the regression class that surfaced as W653 — any future detector that catches Exception without re-raising structural errors fails at PR time.
• W664 — __init__.py package-data drift-guard. CAUGHT A LIVE W643-class bug on first run: roam.languages.extractors was missing its __init__.py, meaning the package was not actually shipped in the wheel despite being referenced via package-data. Pairs with W570/W610 to give the wheel-bundling discipline complete coverage across both YAML/JSON data files AND Python package directories.
• W668 — as_file() callers audit + 4 fixes + drift-guard. Pattern sealed: every importlib.resources.files(...) call that needs a filesystem path now goes through as_file() (4 sites fixed); the drift-guard pins the pattern at lint time. Continues the W554/W535/W610/W624 importlib.resources discipline thread.
• W661 — Fail-loud discipline applied to catalog/detectors production loop. 8 new tests. Production detector loop now classifies NameError/ImportError/AttributeError/TypeError as RuntimeError (structural — re-raise) vs sqlite3.Error (per-detector — swallow+log). Pair with W653 source fix + W662 drift-guard for end-to-end discipline.

Fixed — P0 user-flagged regression batch (W670/W671/W672/W682) + bare-except real bug (W653) + wheel-bundling LIVE bug (W664 finding)

• W670 P0.1 — roam_plan file_path alias regression fixed. Moved _wrap_with_alias_normalization BEFORE the preset filter so the alias normalization layer wraps every tool, not just those that survive the filter. The regression had silently dropped file_path on the roam_plan MCP wrapper when core preset was active. User-flagged P0.
• W671 P0.2 — roam_catalog cold-start auto-handle exemption. New _INLINE_RESPONSE_TOOLS frozenset exempts roam_catalog from _wrap_with_handle_off — the cold-start catalog call now returns inline instead of through a handle that the agent never polled for. Pattern: tools that respond on cold-start with bounded output should bypass the handle pattern. User-flagged P0.
• W672 P0.3 — 8 files synced to live 238/231/224 counts. Auto-derived via dev/build_readme_counts.py --apply (the same path W557 used for version bumps). README + CLAUDE.md + llms-install.md + both mcp-server-card.json copies + server.json + landing-page HTML all carry the live counts. User-flagged P0.
• W682 P0.3-followup — README CLI table: added the evidence-oscal row. Closes the W672 gap audit — the table claimed (all 231) rows but the OSCAL surface was missing. User-flagged P0.
• W653 — REAL BUG: run_all_detectors bare-except now classifies. Pre-fix: bare-except swallowed NameError/ImportError/AttributeError/TypeError as if they were per-detector failures, masking structural bugs (e.g. a typo in a detector module would silently skip the detector and report success). Post-fix: those four classify as RuntimeError (structural — re-raise) while sqlite3.Error keeps the swallow+log behavior. Surfaced by the W662 drift-guard scoping pass.
• W665 — 3 bare-except sites narrowed (allowlist 9 → 6). Continues the W662 drift-guard scoping.
• W677 — 2 more bare-except sites narrowed (allowlist 6 → 4). Continues the W665 scoping.
• W642 — Removed triple-parent fallback from mcp --card handler. -19 LOC. W624 (prior batch) migrated the resolution to importlib.resources.files("roam") / "mcp-server-card.json" with as_file(); the triple-parent parents[3] fallback was dead code after that migration. Continues the importlib.resources discipline thread.

Research/added — Suppression-parser audit (W676) + W646 dogfood refactor + W683/W685 hygiene

• W676 — Suppression parser audit (BAILED — surfaced CRITICAL latent bug). Investigated the suppression parsers to consolidate them; discovered 4 parsers (not 3) with incompatible schemas reading .roam/suppressions.json. Two readers consume the same file with different shape contracts, so a suppression entered for one detector silently does NOT apply to the other reader's detector. W691 in flight to seal this — likely a closed-schema migration to a single canonical parser + a drift-guard. The BAIL is the find: investigate-first discipline saved a fabricated consolidation that would have shipped the bug forward.
• W646 — Refactored _create_extractor from 105 → 17 lines via _LANGUAGE_EXTRACTORS dispatch dict. Eat-our-own-dogfood: W601 (switch-statement smell detector — prior batch) flagged _create_extractor's 23-arm switch as a REAL refactor candidate on roam-code itself; W646 sealed it. First time the smell catalog caught a true positive on roam-code AND the same week's refactor wave sealed it.
• W683 — .gitattributes extended (13 → 49 lines). eol=lf for text files + 26 binary rules. Closes the long-tail "CRLF-on-Windows breaks the wheel-built smoke job" failure class that surfaced intermittently on the W577 wheel-smoke CI.
• W685 — README CLI table header pinned to "(all 231)". Smart 231-canonical choice (matches the canonical count, not the raw 238 command count). Pairs with W672/W682 to keep the README evidence-table column header in lockstep with the auto-derived canonical surface.

Changed — (ADD) W698 consolidation pass

• W698 — CHANGELOG/HANDOVER/BACKLOG/SESSION-SNAPSHOT refresh for W642-W685 batch. Docs-only; hash-stability mandate held trivially. 51/51 doc-consistency + schema-migration tests pass.

> Pattern-3a vocabulary cluster GENUINELY STRUCTURALLY CLOSED across ALL THREE rank axes (severity + confidence + risk) + smell-detector catalog reached 20 detectors (was 15) + _wrap_with_alias_normalization refactor+dedup chain + cross-detector empty-corpus smoke (W607 / W624 / W631 / W601 / W602 / W640 / W605 / W648 / W639 / W636 batch, 2026-05-15). > Nine-completion batch folded in behind the W635-CONSOLIDATE consolidation. The headline is Pattern-3a GENUINELY closed end-to-end: W631 introduces the third canonical axis src/roam/output/risk.py::risk_rank() and migrates 2 sites (cmd_migration_plan + cmd_path_coverage), pairing with W564 severity-rank + W596 confidence-rank to canonicalize ALL THREE rank axes (severity + confidence + risk); W648 AST audit returned ZERO slipped rank tables across the entire src tree — Pattern-3a is structurally closed for real, not just-in-name; W640 folded cmd_alerts._LEVEL_ORDER into severity_rank() via -severity_rank(lowercase) and broadened the drift-guard regex /sever/ → /sever|level_order/. Smell catalog reached 20 detectors (was 15 at session start): W370c 5-smell expansion COMPLETE — W601 (switch-statement, 7 findings; surfaced REAL refactor candidate _create_extractor 23-arm switch), W602 (temporal-coupling, 10 findings; surfaced cli↔_run_roam_inprocess 34-commit top coupling), W603 (magic-numbers), W604 (boolean-parameter), W605 (comment-density TODO/FIXME/XXX/HACK; roam-code CLEAN at max 0.49% rate). LAW-4 anchor sets bumped 92→93 / 109→110 to accommodate comment-density terminals. _wrap_with_alias_normalization refactor + dedup chain complete: W607 decomposed the 130-line _wrap_with_alias_normalization into 3 helpers (_collect_alias_candidates, _build_merged_signature, _build_merged_annotations; 130→50 lines + 7 unit tests; 2960 focused tests pass); W636 collapsed the sync/async wrapper closure duplication via shared _prepare_kwargs helper + branched closure (33→28 lines + duplicate-body anti-pattern eliminated). Pairs with W595 (param-ordering seal) + W606 (canonical-positional collision lint) to give the wrapper its end-to-end discipline. Cross-detector empty-corpus smoke (W639) guards 54 detectors (20 smells + 34 algo + 2 floor counts; 56+115+17+31 = 219 tests) against silent import errors after concurrent merges — catches the W601/W602-style regression class at PR time. W624 migrated the mcp --card handler at mcp_server.py:14593-14624 to importlib.resources.files("roam") / "mcp-server-card.json" with as_file() — completes the importlib.resources discipline thread (10+31+140 tests pass). Hash-stability 31/31 byte-identical held across every source wave.

Added — Risk-rank canonical helper + 5 new smell detectors + cross-detector empty-corpus smoke (W631/W601/W602/W603/W604/W605/W639 batch)

• W631 — Third Pattern-3a axis CANONICALIZED. New src/roam/output/risk.py::risk_rank() helper; 2 sites migrated (cmd_migration_plan + cmd_path_coverage). Pattern-3a STRUCTURALLY CLOSED ACROSS ALL THREE AXES (severity W547+W564 + confidence W596 + risk W631). 131 tests pass. Direct fulfillment of the W635-batch carry-forward.
• W601 — switch-statement smell detector. 7 findings on roam-code itself, surfaced a REAL refactor candidate: _create_extractor 23-arm switch. Continues the W370c 5-smell expansion thread.
• W602 — temporal-coupling smell detector. 10 findings; top coupling is cli ↔ _run_roam_inprocess at 34 commits. Surfaces the recurring "two files change together but live in different modules" anti-pattern. Continues W370c.
• W603 — magic-numbers smell detector. Continues W370c.
• W604 — boolean-parameter smell detector. Continues W370c.
• W605 — comment-density smell detector (TODO / FIXME / XXX / HACK). 20th detector ships; roam-code CLEAN (max rate 0.49%). 173 tests pass. Closes W370c 5-smell expansion (W601/W602/W603/W604/W605 all shipped). LAW-4 concrete-noun anchor sets bumped 92 → 93 (formatter) and 109 → 110 (test) to accommodate the comment-density finding terminals.
• W639 — Cross-detector empty-corpus smoke test. Guards 54 detectors (20 smells + 34 algo + 2 floor counts) against silent import errors after concurrent merges. 56 + 115 + 17 + 31 = 219 tests. Catches the W601 / W602-style concurrent-merge import-error regression class at PR time — a class that previously surfaced only on the next dogfood pass.
• W601 + W602 bundle — 17 → 19 detector count milestone. 165 tests pass on the bundle's structural pieces. W605 carries the count from 19 → 20.

Changed — _wrap_with_alias_normalization refactor + dedup + importlib.resources migration + alerts level_order fold (W607/W636/W624/W640/W648 batch)

• W607 — Decomposed _wrap_with_alias_normalization into 3 helpers. _collect_alias_candidates, _build_merged_signature, _build_merged_annotations. 130 → 50 lines. 7 unit tests + 2960 focused tests pass. Continues the W595 (param-ordering seal) + W606 (canonical-positional collision lint) refactor thread on the wrapper.
• W636 — Sync/async wrapper closure duplication collapsed. New shared _prepare_kwargs helper + branched closure pattern. 33 → 28 lines and duplicate-body anti-pattern eliminated. 40 tests pass. Pairs with W607 to give the wrapper its final shape.
• W624 — mcp --card handler migrated to importlib.resources. mcp_server.py:14593-14624 now resolves via importlib.resources.files("roam") / "mcp-server-card.json" with as_file(). 10 + 31 + 140 tests pass. Continues the W554 / W535 / W610 importlib.resources discipline thread.
• W640 — cmd_alerts._LEVEL_ORDER folded into severity_rank(). Sort key now uses -severity_rank(lowercase) instead of a private rank table. Drift-guard regex broadened /sever/ → /sever|level_order/ so the canonicalization is enforced for any future _LEVEL_ORDER clones. 121 tests pass. One of the W648 audit's findings, sealed in the same wave.
• W648 — AST audit for slipped rank tables. ZERO slipped. Audited the entire src tree for inline rank tables that would bypass the canonical severity_rank() / confidence_level_rank() / risk_rank() helpers. Result: zero slipped. Pattern-3a is GENUINELY structurally closed — the audit confirms the structural-close claim is not just-in-name. 47/47 + 31/31 tests pass.

> Pattern-3a vocabulary cluster STRUCTURALLY CLOSED across BOTH rank axes (severity + confidence) + smell-detector catalog reached ZERO placeholder stubs + wheel-bundling discipline complete + fragile-path sweep + AST drift-guards across the board (W596 / W594 / W588 / W577 / W570 / W564 / W515 / W370c batch, 2026-05-15). > Sixteen-completion batch folded in behind the W600-CONSOLIDATE consolidation. The headline is structural across BOTH rank axes: W596 completes Pattern-3a confidence-rank consolidation by migrating 15 sites to the canonical src/roam/output/confidence.py::confidence_level_rank() helper (561 tests pass), pairing with W564's prior 10-site severity-rank migration to close the Pattern-3a vocabulary cluster end-to-end. Combined with W547 (severity vocab) + W518 (control-mapping vocab) + W512 (edge-kinds) + W565+W566 (severity helpers), drift-guard discipline now canonicalizes through 6 modules + 6 AST lint suites — every Pattern-3a vocabulary cluster surfaced in the dogfood corpus flows through canonical modules with AST drift-guards. Third rank axis (risk) flagged as W631 follow-up. Smell detector catalog reached ZERO placeholder stubs (W370c): shipped 2 detectors (refused-bequest 2 findings + primitive-obsession 144 findings) and scoped the remaining stubs into 5 W370c-followup waves (W601-W605) for new smell kinds. Fragile-path harness gotcha closed end-to-end: W587 (10 sites) + W594 (18 sites, 47 → 29 remaining) swept 28 of 57 fragile-path test sites to the canonical tests/_helpers/repo_root.py helper; W588 added an AST drift-guard for the Path(__file__).parents[N] pattern with fail-loud _PRE_W594_PENDING allowlist (47 entries — corrected upward from the 27 estimate by the W588 inventory pass); W606 added an AST lint for canonical-positional collision catching the pre-W595 crash class at PR time (4 new tests). Wheel-bundling discipline COMPLETE: W554 customer SHIPPING BUG fixed + W570 drift-guard + W577 CI wheel-smoke job (3 steps: build wheel + install fresh venv + run drift-guard from /tmp) + W610 extended to taint_rules + languages.extractors + mcp-server-card (3 new test classes, 6 new tests) — closes prior 2 silent-empty bugs (12.12.1 taint rules + 12.12.2 Jenkinsfile) across 5 package-data surfaces. Pattern 1 variant D family CLOSED at the CLI boundary (W573 NO-OP investigation confirmed only 1 production call site for ChangeEvidence.from_canonical_json exists). Leasing-system parity completed: W447 + W448 (bundled) added the pr-replay info marker on missing leases dir + read_lease(warnings_out=...) kwarg. Severity helpers landed: W565 + W566 (bundled) added severity_to_confidence_level() + severity_breakdown() helpers (5 call-sites migrated, 248 tests). Drift-guard parsing seal: W515 parses python-version from the live workflow before drift compare, sealing the false-positive class on CI version bumps (139 tests). Doc sweep: W569 swept 9 stale templates/audit-report/ path refs across 8 src/dev files + 1 test docstring (111 tests). Small cleanups: W591-bundle W584 / W497 / W500 bailed as already-done; W501 audit comments added to 4 test files. W573 NO-OP investigation: only 1 production call site for ChangeEvidence.from_canonical_json exists — Pattern 1 variant D family fully sealed at CLI boundary. Hash-stability 31/31 byte-identical held across every source wave.

Research/added — Pattern-3a confidence-rank canonicalization + smell detectors + AST drift-guards + wheel-smoke CI (W370c/W515/W564/W577/W588/W596/W606/W610 batch)

• W596 — MASSIVE Pattern-3a confidence-rank consolidation. 15 sites migrated to the canonical src/roam/output/confidence.py::confidence_level_rank() helper. 561 tests pass. 31/31 hash-stability byte-identical held. Closes the Pattern-3a vocabulary cluster across BOTH rank axes — the pair to W564 severity-rank: every Pattern-3a vocabulary surface surfaced in the dogfood corpus now flows through canonical modules with AST drift-guards. Third rank axis (risk) flagged as W631 follow-up.
• W370c — Smell detector catalog reached ZERO placeholder stubs. Scoped the W368 BEHIND-list smell stubs + shipped 2 detectors: refused-bequest (2 findings) + primitive-obsession (144 findings). Catalog now has ZERO placeholder detectors. 5 W370c-followup waves (W601-W605) queued for new smell kinds (first 2 in flight). Closes the catalog-cleanup thread that has been carry-forward since W368.
• W588 — AST drift-guard for fragile-path Path(__file__).parents[N] pattern. Fail-loud with _PRE_W594_PENDING allowlist (47 entries — corrected upward from the 27 estimate by the W588 inventory). Companion to the W587 + W594 sweep below. Pairs with W606 canonical-positional collision lint to give the fragile-path harness gotcha end-to-end lint coverage.
• W606 — AST lint for canonical-positional collision. 4 new tests catching the pre-W595 crash class at PR time. Closes the latent breakage class that the W587 fragile-path sweep surfaced (_wrap_with_alias_normalization param-ordering — W595 sealed the source bug; W606 ensures regression catches at lint time).
• W577 — Wheel-built CI smoke job added to roam-ci.yml. 3 steps: build wheel + install into fresh venv + run drift-guard from /tmp. Pairs with W570 (drift-guard) + W610 (extension to 5 package-data surfaces) to close the customer-facing shipping-bug class structurally — pip install roam-code users can no longer hit a "feature works in src but broken on wheel install" surface without CI catching it.
• W610 — Wheel drift-guard extended to 3 more package-data surfaces. Adds taint_rules + languages.extractors + mcp-server-card to the W570 pin. 3 new test classes + 6 new tests. Closes 5 package-data surfaces end-to-end (W570 pinned 2; W610 pinned 3 more). Sealing the prior 2 silent-empty bugs in the wheel (12.12.1 taint rules + 12.12.2 Jenkinsfile).
• W515 — Drift-guard parses python-version from the live workflow before drift compare. False-positive class on CI version bumps sealed: the lint no longer flags routine version-string bumps as drift. 139 tests pass. Closes the long-tail false-positive class that surfaced on every Python-3.X minor-version bump.
• W564 — MASSIVE Pattern-3a severity-rank consolidation (carry from W591 batch). 10 sites migrated to canonical severity_rank(). 460 + 31 tests pass. Listed here as the structural pair to W596 in the BOTH-rank-axes structural close-out chain.

Fixed — Fragile-path sweep continues + Pattern 1 variant D CLI boundary closed + leasing parity (W447/W448/W573/W587/W594 batch)

• W594 — 18 fragile-path test sites migrated to tests/_helpers/repo_root.py (47 → 29 remaining). Continues the W587 sweep (which migrated the first 10 highest-noise sites). W608 priority pair (W512+W547 drift-guard templates) included. 230 tests pass. W588 drift-guard now fail-loud on any regression — the remaining 29 sites are in the _PRE_W594_PENDING allowlist with explicit migration tasks tracked as W612-W613.
• W587 — 10 fragile-path test sites migrated (carry from W591 batch). Closes the worktree-vs-main-tree visibility gotcha (W567) for the 10 highest-noise sites; pairs with W594 (next-18) + W588 (drift-guard) + W606 (canonical-positional collision lint) to give the fragile-path harness end-to-end lint discipline.
• W447 + W448 (bundled) — Leasing-system Pattern-2 always-emit discipline complete (carry from W591 batch). W447 added the pr-replay info marker on missing leases dir under migration / autonomous_pr modes. W448 added the roam.leases.store.read_lease(warnings_out=...) kwarg.
• W573 — Pattern 1 variant D family CLOSED at the CLI boundary (NO-OP investigation, carry from W591 batch). Only 1 production call site for ChangeEvidence.from_canonical_json* exists. Pattern 1 variant D family fully sealed at the CLI boundary.

Changed — Severity helpers + doc sweep + small cleanups (W565/W566/W569/W591-bundle batch — carry from W591 batch)

• W565 + W566 (bundled) — Severity helpers landed in _severity.py. New severity_to_confidence_level() + severity_breakdown() helpers. 5 call-sites migrated. 248 tests pass.
• W569 — Doc sweep: 9 stale templates/audit-report/ path refs swept across 8 src/dev files + 1 test docstring. 111 tests pass.
• W591-bundle — Small cleanups. W584 / W497 / W500 bailed as already-done; W501 audit comments added to 4 test files. 81 tests pass.

> Pattern-3a severity-rank consolidation STRUCTURALLY CLOSED + fragile-path sweep + leasing parity + git-helper consolidation + Pattern 1 variant D CLI-boundary close (W540-W591 batch, 2026-05-15). > Nine-completion batch folded in behind the W578-CONSOLIDATE consolidation. The headline is structural: W564 completes the Pattern-3a severity-rank consolidation by migrating 10 sites to the canonical severity_rank() helper alongside W512 (edge-kinds) + W518 (control-mapping vocab) + W547 (severity vocab) + W565+W566 (severity helpers) — every Pattern-3a vocabulary cluster surfaced in the dogfood corpus now flows through canonical modules with AST drift-guards. 14 confidence-rank tables flagged as the next Pattern-3a target (W596 queued). 460 + 31 tests pass. Pattern 1 variant D CLI boundary CLOSED: W573 investigation confirmed only 1 production call site for ChangeEvidence.from_canonical_json* exists (the one W561 already migrated) — the variant D family is fully sealed at the CLI boundary. Leasing-system parity completed: W447 + W448 (bundled) added the pr-replay info marker on missing leases dir under migration / autonomous_pr modes + the roam.leases.store.read_lease(warnings_out=...) kwarg — Pattern-2 always-emit discipline now covers list_leases (W425) + read_lease (W448) + pr-replay info-marker (W447) end-to-end. 137 + 31 tests pass. Git-helper subprocess discipline: W540 consolidated _git_fingerprint + _git_commit_sha helpers; pr-bundle init now shells out to git rev-parse HEAD ONCE instead of TWICE per invocation. 105 + 31 tests pass. Severity helpers landed: W565 + W566 (bundled) added severity_to_confidence_level() + severity_breakdown() to _severity.py with 5 call-sites migrated. 248 tests pass. Fragile-path sweep (W587): 10 test sites migrated to the new tests/_helpers/repo_root.py helper — 37 → 27 fragile-path sites remain (W594 queued for the remainder). Surfaced a real bug: _wrap_with_alias_normalization param-ordering breaks test_surface_consistency (W595 in flight). Small cleanups (W591-bundle): W584 / W497 / W500 bailed as already-done; W501 audit comments added to 4 test files. 81 tests pass. Doc sweep (W569): 9 stale templates/audit-report/ path refs swept across 8 src/dev files + 1 test docstring + 1 fixture-regen command. 111 tests pass. Hash-stability 31/31 byte-identical held across every source wave.

Changed — Pattern-3a severity-rank canonicalization + severity helpers + git-helper consolidation (W540/W564/W565/W566 batch)

• W564 — MASSIVE Pattern-3a severity-rank consolidation. 10 sites migrated to the canonical severity_rank() helper. 460 tests pass. 31/31 hash-stability byte-identical held. 14 confidence-rank tables flagged as the next Pattern-3a target (W596 queued). Continues the structural Pattern-3a close-out chain: W512 (edge-kinds) + W518 (control-mapping vocab) + W547 (severity vocab) + W564 (severity-rank) + W565+W566 (severity helpers) — every cluster surfaced in the dogfood corpus now flows through canonical modules with AST drift-guards.
• W565 + W566 (bundled) — Severity helpers landed in _severity.py. New severity_to_confidence_level() + severity_breakdown() helpers. 5 call-sites migrated. 248 tests pass. Pair with W547/W548 to give every consumer one canonical entry-point per severity-derived computation.
• W540 — Consolidated _git_fingerprint + _git_commit_sha helpers. pr-bundle init now shells out to git rev-parse HEAD ONCE per invocation instead of TWICE — closes the subprocess-discipline gap surfaced by the W521 producer-side commit_sha stamping work. 105 tests pass. 31/31 hash-stability byte-identical held.

Fixed — Leasing parity + Pattern 1 variant D CLI boundary closed (W447/W448/W573 batch)

• W447 + W448 (bundled) — Leasing-system Pattern-2 always-emit discipline complete. W447 added the pr-replay info marker on missing leases dir under migration / autonomous_pr modes — explicit state: "leases_not_initialized" rather than silent SAFE. W448 added the roam.leases.store.read_lease(warnings_out=...) kwarg — pairs with W425's list_leases(warnings_out=...) to give every lease read path the same always-emit surface. 137 + 31 tests pass.
• W573 — Pattern 1 variant D family CLOSED at the CLI boundary (NO-OP investigation). Confirmed only 1 production call site for ChangeEvidence.from_canonical_json* exists (the one W561 already migrated to surface dropped_enum_rows + partial_success). The variant D class — "silent success on degraded resolution" — is fully sealed at the CLI boundary as of this investigation. No further migration work needed; class structurally closed.

Changed — Fragile-path sweep + small cleanups + doc sweep (W569/W587/W591 batch)

• W587 — 10 fragile-path test sites migrated to tests/_helpers/repo_root.py. Closes the worktree-vs-main-tree visibility gotcha (W567) for the 10 highest-noise sites; 37 → 27 fragile-path sites remaining (W594 queued). Surfaced a real bug: _wrap_with_alias_normalization param-ordering breaks test_surface_consistency (W595 in flight) — exactly the kind of latent breakage the worktree-vs-main-tree visibility gap was masking.
• W591-bundle — Small cleanups. W584 / W497 / W500 bailed as already-done (investigate-first discipline saved fabricating work); W501 audit comments added to 4 test files. 81 tests pass.
• W569 — Doc sweep: 9 stale templates/audit-report/ path refs swept. Across 8 src/dev files + 1 test docstring + 1 fixture-regen command. 111 tests pass. Closes the long-tail doc-drift class surfaced by the W554 move (control-mapping.yaml moved into src/roam/templates/audit_report/ — referrers needed to follow).

> SHIPPING BUG FIXED + Pattern 1 variant D disclosure + canonical severity vocab + ChangeEvidence round-trip pipeline + OSCAL persistent artifacts + package-data drift-guard (W520-W570, 2026-05-15). > Ten-completion batch folded in behind the W549 consolidation. The headline is a customer-facing shipping bug fix: W554 moved templates/audit-report/control-mapping.yaml into src/roam/templates/audit_report/ + added the pyproject.toml package-data entry — pip install roam-code users could not previously run roam ci-setup --with-oscal or roam evidence-oscal against their own projects because the control-mapping YAML was not bundled in the wheel. Lookup migrated to importlib.resources. Verified end-to-end via fresh tmp venv wheel install (109 tests pass). Pattern 1 variant D dropped_enum_rows disclosure lands across the AR envelope: W534 introduced ChangeEvidence.from_canonical_json(text, *, strict=False) with closed-enum validation — 31 golden fixtures round-trip BYTE-IDENTICAL with content hashes preserved (forgiving projection mode); W561 added from_canonical_json_with_drops() classmethod that surfaces dropped enum rows + partial_success: true on the envelope (LAW-4 anchored on rows terminal); W559 wired from_canonical_json into the cmd_evidence_oscal AR path with a --strict flag (hybrid Mapping|ChangeEvidence signature). Forgiving-projection AND fail-loud discipline now both available end-to-end (W465 golden fixture stays byte-identical). Canonical severity vocabulary in src/roam/output/_severity.py (W547 + W548 bundled): SEVERITY_LEVELS / SEVERITY_ALIASES / normalize_severity / to_sarif_level / validate_severity + AST drift-guard — closes the Pattern 3a severity-vocabulary divergence across SARIF emitters. OSCAL persistent artifacts (W535): roam ci-setup --with-oscal now materializes .roam/oscal/control-mapping.json + stub-assessment-plan.json with deterministic UUIDv5 + SHA-256-seeded timestamps — the FedRAMP continuous-assessment evidence pattern. SLSA SRC-L3 commit_sha chain CLOSED: W520 added the cga-sibling emit_cga_vsa_sibling commit_sha fallback — belt-and-suspenders complement to W509 — completing the producer-W521 + collector-W509 + cga-sibling-W520 three-path chain (all three fall back to git rev-parse HEAD). Package-data wheel-bundling discipline: W570 added tests/test_package_data_wheel_drift.py drift-guard pinning roam.templates.audit_report + roam.templates.ci package-data entries. Closes the recurring "feature works in src but broken on pip install" surface (the W554-class bug). Version-skew + hash-stability hygiene: W557 rolled server.json + mcp-server-card.json 12.50→13.0 via dev/build_readme_counts.py --apply; W563 normalizes auto-derived fields before hashing in the card-hash test so count/version bumps stay invisible while preserving the R17 tampering guard for other fields. Hash-stability 31/31 byte-identical held across every source wave.

Research/added — ChangeEvidence round-trip + canonical severity + OSCAL persistence + cga commit_sha (W520/W534/W535/W547/W548/W559/W561/W563/W570 batch)

• **W534 — ChangeEvidence.from_canonical_json(text, *, strict=False). Closed-enum validation; 31 golden fixtures round-trip BYTE-IDENTICAL** with content hashes preserved. Forgiving-projection mode by default; strict=True raises on unknown enum values. The structural answer to "how do consumers safely round-trip an AR envelope without breaking the byte-identical hash discipline" — fully passes both halves of the contract.
• W535 — roam ci-setup --with-oscal persistent artifacts. Materializes .roam/oscal/control-mapping.json + .roam/oscal/stub-assessment-plan.json with deterministic UUIDv5 (namespace-pinned) + SHA-256-seeded timestamps so re-runs produce byte-identical outputs. 21 + 15 + 16 + 31 tests pass. Pattern: FedRAMP continuous-assessment requires durable artifacts on disk, not ephemeral envelopes — --with-oscal is the bootstrap pair to the W465 roam evidence-oscal runtime emitter.
• W547 + W548 (bundled) — Canonical src/roam/output/_severity.py module. Single source of truth for SEVERITY_LEVELS / SEVERITY_ALIASES / normalize_severity / to_sarif_level / validate_severity. AST drift-guard pins the closed enumeration at construction time. 89 tests pass. Closes Pattern 3a (vocabulary divergence) on severity across SARIF emitters — every emitter now resolves through the same canonical helpers.
• W559 — Wired ChangeEvidence.from_canonical_json into cmd_evidence_oscal AR path with --strict flag. Hybrid Mapping|ChangeEvidence signature so callers can pass either a raw mapping (legacy / forgiving) or a parsed ChangeEvidence instance (typed / strict). W465 golden fixture stays byte-identical. 116 tests pass.
• W561 — Pattern 1 variant D dropped_enum_rows + partial_success disclosure on AR envelope. New from_canonical_json_with_drops() classmethod returns (evidence, dropped_rows) so the consumer surface can disclose silent enum-drop side effects rather than silently projecting them away. LAW-4 anchored on the rows terminal. 107 + 176 tests pass. Direct application of the dogfood synthesis "silent success on degraded resolution" guard.
• W563 — Card-hash test normalizes auto-derived fields before hashing. Hybrid A+B: count/version bumps are invisible to the hash; R17 tampering guard preserved for every other field. Makes routine version/count bumps a no-op against the card-hash drift-guard without weakening the integrity claim. 3 + 10 + 5 + 31 tests pass.
• W520 — emit_cga_vsa_sibling commit_sha fallback (belt-and-suspenders complement to W509). Adds git rev-parse HEAD fallback to the cga sibling emit path, completing the SLSA SRC-L3 commit_sha chain end-to-end: producer W521 stamps at pr-bundle init, collector W509 falls back at pr-bundle emit, and cga sibling W520 falls back on the cga-emit-time path. All three paths now carry commit_sha through the no-collect path — the W498-surfaced drift class is fully sealed across pr-bundle AND cga surfaces.
• W570 — tests/test_package_data_wheel_drift.py drift-guard. Pins roam.templates.audit_report + roam.templates.ci package-data entries in pyproject.toml. 4 + 24 + 15 + 31 tests pass. Structural answer to the recurring "feature works in src but broken on pip install" failure mode that produced W554 — drift on package-data entries is now caught at lint time.

Fixed — SHIPPING BUG closed (W554) + version skew (W557)

• W554 — SHIPPING BUG: templates/audit-report/control-mapping.yaml MOVED into src/roam/templates/audit_report/ + pyproject.toml package-data entry added. pip install roam-code users could not previously run roam ci-setup --with-oscal or roam evidence-oscal against their own projects because the control-mapping YAML was not bundled in the wheel — the surface worked end-to-end in the src tree and silently broke after pip install. Lookup migrated to importlib.resources. Verified end-to-end via fresh tmp venv wheel install (built wheel, installed into a throwaway venv, ran roam ci-setup --with-oscal + roam evidence-oscal against the venv-installed binary). 109 tests pass. Pairs with W570 drift-guard above to prevent regression.
• W557 — Version skew fix on server.json + mcp-server-card.json. Rolled 12.50 → 13.0 via dev/build_readme_counts.py --apply (the auto-derived path; manual edits would have re-drifted). 60 tests pass. The card-hash R17 drift-guard rolled forward via W563's normalize-before-hash change so the bump landed invisibly.

> W493 BUG FAMILY STRUCTURALLY CLOSED + THREE more silent no-ops sealed + OSCAL pipeline end-to-end + OWASP labels integrity + SLSA SRC-L3 commit_sha parity (W506-W533, 2026-05-15). > Ten-wave batch closed behind the W516 docs consolidation. The headline is structural: W512 introduces src/roam/db/edge_kinds.py + a 16-test drift-guard lint that migrates 12 read-sites to canonical helpers and structurally seals the W493/W499/W511/W524 edge-kind bug family — future inline kind IN queries fail at lint time. Three more long-latent silent no-ops sealed this batch: (1) W511 fixed side_effects.py:497 edge-kind union (production impact 13/14,949 → 14,949/14,949 edges matched — the FOURTH silent no-op in the W493 family); (2) W524-bundle hunt found 7,534 missing import edges in cmd_hover.py (the largest single edge-kind no-op in the family by 3 orders of magnitude — hover output had been blind to imports since launch), plus +13 references in cmd_risk.py and defensive plumbing in cmd_patterns.py; (3) W531 caught SARIF severity=error silently downgrading to "note" since launch — GitHub Code Scanning + Microsoft Defender were not flagging taint findings as errors for any consumer that ingested roam SARIF, ever. OSCAL pipeline fully shipped end-to-end: W465 added Assessment Results emission via roam evidence-oscal --kind assessment-results (with auto-synthesized stub Assessment Plan per the FedRAMP continuous-assessment pattern). With W464 already in flight, roam evidence-oscal now covers both v1.2 models. Claim-integrity batch on OWASP labels: W533-bundle (W530+W531+W532) corrected the OWASP A05 → A03 mislabel on java_sqli + python_ssti and brought owasp_top10 coverage from 3/22 → 22/22 rules, plumbed via W492/W453 into TaintRule / TaintFinding / findings.evidence_json / SARIF tags[]. SLSA SRC-L3 commit_sha parity completed: W509 added the emit-time git rev-parse HEAD fallback (restoring cga sibling parity surfaced by W498), and W521 stamped commit_sha producer-side at pr-bundle init so the W509 fallback becomes belt-and-suspenders. Framework-vocab consolidation: W518 collapsed scattered allowlists into src/roam/evidence/control_mapping_vocab.py (9 framework slugs + 9 titles + 3 pass-conditions + 7 surfaces) with drift-guard. SLSA control-map entries shipped: W506 landed the 3 missing SRC-L2/L3 entries + iso_42001 → iso_iec_42001 rename across 5 files in lockstep — claim-integrity hygiene now matches the W451/W471/W472 SRC-L3 pipeline. Hash-stability 31/31 byte-identical held across every source wave.

Fixed — THREE long-latent silent no-ops + W493 family structurally sealed (W506-W533 batch)

• W512 — STRUCTURAL CLOSE of the W493/W499/W511/W524 edge-kind bug family. New src/roam/db/edge_kinds.py closed-enum module with canonical helpers; 12 read-sites migrated to call the helpers instead of inlining kind IN (...) literal-string tuples. 16-test drift-guard lint added — future inline kind IN queries against the edges table fail at lint time. This is the structural answer to the same edge-kind class that produced W493 (taint DFS no-op), W499 (impact gate no-op), W511 (effects propagation no-op), and W524 (cmd_hover 7k missing imports). 365 tests pass.
• W511 — side_effects.py:497 edge-kind union (CRITICAL CORRECTNESS). The FOURTH silent no-op in the W493 family. effects_propagation was matching 13 / 14,949 edges pre-fix (0.087% coverage); post-fix matches 14,949 / 14,949 edges (100%). Side-effect classification had been computed against a near-empty subset of the call graph since the edge-kind canonicals diverged. Caught by the same dogfood pattern that found W493 + W499.
• W524-bundle — Phantom edge-kind hunt + 3 broken sites fixed. Audited the rest of the codebase for the W493/W499/W511 class. Three sites repaired: cmd_risk.py +13 references, cmd_hover.py +7,534 import edges (massive missing — hover output had been blind to imports since launch, the largest single edge-kind no-op in the family), and defensive plumbing in cmd_patterns.py. 202 tests pass.
• W531 — SARIF severity=error silently downgraded to "note" since launch. Discovered via the W533-bundle audit: SARIF emission was setting level="error" correctly but the downstream serializer was downgrading on the wire to "note". GitHub Code Scanning + Microsoft Defender for DevOps + every SARIF-ingesting tool was NOT flagging roam taint findings as errors since the SARIF feature launched. Fix restores level="error" end-to-end so the critical-severity surface fires correctly in CI.
• W530 — OWASP A05 → A03 mislabel on java_sqli + python_ssti (CLAIM INTEGRITY). Both rules were stamped A05:2021 (Security Misconfiguration) when they should have been A03:2021 (Injection). Audit + correction landed alongside W531 + W532 in the W533-bundle.
• W532 — owasp_top10 coverage 3/22 → 22/22 rules. Only 3 of 22 taint rules carried owasp_top10 stamps pre-fix; all 22 now correctly carry the annotation, surfaced via SARIF tags[] (W453 plumbing) and findings.evidence_json (W492 plumbing).
• W509 — pr-bundle emit commit_sha fallback via git rev-parse HEAD. Sealed the W498-surfaced drift: pr-bundle was dropping commit_sha on --no-auto-collect while cga emit fell back to git. The fix restores SRC-L3 commit-anchored provenance parity with the cga path. (W521 then made it belt-and-suspenders by stamping commit_sha producer-side at bundle init — see Added section below.)

Added — OSCAL Assessment Results + OWASP plumbing + framework-vocab module + SLSA entries + producer-side commit_sha (W506-W533 batch)

• W465 — OSCAL v1.2 Assessment Results emission. roam evidence-oscal --kind assessment-results now emits v1.2 Assessment Results JSON; a stub Assessment Plan is auto-synthesized when no upstream plan exists (FedRAMP continuous-assessment pattern). Combined with the in-flight W464 Control Mapping emitter, roam evidence-oscal covers both OSCAL v1.2 models end-to-end. 81 tests pass.
• W492 + W453 — owasp_top10 wired through the taint pipeline end-to-end. Loaded into TaintRule + TaintFinding dataclasses, persisted to findings.evidence_json, and plumbed to SARIF tags[]. 207 tests pass. Pairs with the W533-bundle claim-integrity fixes above.
• W518 — Framework-vocab allowlist consolidation. New src/roam/evidence/control_mapping_vocab.py collapses the scattered framework-vocab allowlists into a single module: 9 framework slugs + 9 titles + 3 pass-conditions + 7 surfaces, with a drift-guard test pinning the closed enumerations. Same shape as the W332 / W282 / W211 / W505-bundle vocabulary-freeze discipline.
• W506 — SLSA SRC-L2/L3 control-mapping entries. 3 new entries landed in templates/audit-report/control-map.yml alongside W428's NIST AI 600-1 + SP 800-218A additions. iso_42001 → iso_iec_42001 rename propagated in lockstep across 5 files so the W518 framework-vocab drift-guard stays green. Claim-integrity hygiene now matches the W451/W471/W472 SRC-L3 substrate.
• W521 — pr-bundle init producer-side commit_sha stamping. Records commit_sha at bundle-init time (single git rev-parse HEAD call). The W509 fallback at emit time becomes belt-and-suspenders — bundles created on a no-collect path now carry commit_sha from the moment of creation. 127 tests pass.

> TWO long-latent silent no-ops sealed + SLSA SRC-L3 evidence-pipeline polish + closed-enum lints + taint trio closure (W375-W515, 2026-05-15). > The wave between the W491 consolidation and this one shipped twelve > threads in parallel. The headline is two critical-correctness > fixes that landed back-to-back: (1) W493 fixed > propagate_taint's kind='calls' query against writers that emit > kind='call' — the taint DFS had been a NO-OP since inception, all > 76 production findings stuck at chain_length=1. Three read-side > sites repaired (taint.py:491, cmd_dead.py:1565, dataflow.py:329); > 4 stale tests that asserted the no-op behavior flipped to assert > the real contract; 31/31 byte-identical golden hashes hold, 292 > tests pass; W441's 607-finding projection now stands for the > production roam-code corpus. (2) W499 fixed > critique/checks.py:399 — the impact gate was matching 0/14,949 > caller edges (COMPLETE NO-OP); post-fix surfaces 5 high-severity > findings on roam-code itself. PRs touching open_db / > json_envelope / to_json / invoke_cli / path now correctly > exit-5 in --ci mode. (3) W375 closed the W372-research first-ship > taint-rule trio (after W373 python-ssti + W374 java-sqli): > java-deserialization rule pack at > src/roam/security/taint_rules/java_deserialization.yaml > (T-X04 / CWE-502 / A08:2021; 15 sources / 12 sinks / 13 sanitizers, > qualified_only: true). (4) W486 extracted the shared > src/roam/attest/emit_vsa.py helper (339 lines); cmd_pr_bundle > and cmd_cga collapse to 9-line + 24-line delegations > respectively. 143/143 tests pass. (5) W498 added the > end-to-end VSA parity test in tests/test_attest_vsa.py:661+ > (TestVsaCliParity) — found real drift: pr-bundle drops > commit_sha when --no-auto-collect; cga falls back to > git rev-parse HEAD. Spawned W509 fix (now in flight). > (6) W428 shipped the 5 W360-research crosswalk YAML entries > (NIST AI 600-1 + SP 800-218A): AI600_VALUE_CHAIN_PROVENANCE, > AI600_STOP_BUILD_AUTHORITY, SSDF218A_CODE_PROVENANCE, > SSDF218A_CODE_REVIEW_AI_OUTPUT, SSDF218A_DEVELOPER_AUTHORIZATION. > CAISI held to H2 2026. W506 in flight to add the missing SLSA > entries — claim-integrity hygiene per the agentic-assurance > "supports evidence for" lint. (7) W505-bundle shipped 3 > closed-enum lints (W502 source_framework / W503 pass_condition / > W504 surface); 19+31 tests pass. (8) W482 added a roam > doctor advisory check that compares the local > .github/workflows/roam.yml against the canonical CI template; > chose advisory-check over a standalone command for low-friction > surfacing. Real-world signal: roam-code's own roam.yml has > drifted from template (26 vs 28 lines) — surfaced on the > dogfooded doctor run. 9 new tests + 137/137 focused pass. > (9) W485 verdict was MEASUREMENT DRIFT, not regression — > the W408 baseline was a 17k-symbol corpus; current roam-code is > 23.6k symbols / 29.9k edges / 3.8k files (+39% / +76% / 7x). > Effects_taint scaled 67.6s → 87.4s; relative dominance held > 48% → 50.5%. (10) W488 auditing pass: the rest of the > test_taint_*.py corpus for stale bare-name assertions came up > CLEAN — W479 caught the only offender; 128+31 tests pass. > (11) W441 BAILED with a high-impact find — it was the > investigation that surfaced the W493 kind='calls' vs kind='call' > typo (real wallclock when fed correct data: 0.06s; W433-research's > 35s prediction was based on stale code). Spawned the critical > W493 fix. (12) W491-CONSOLIDATE — itself (folded inline in > the previous Unreleased entry).

Added — taint trio close + SLSA polish + crosswalk + closed-enum lints + advisory check (W515 batch)

• W375 — OWASP taint rule pack v1 java-deserialization. New src/roam/security/taint_rules/java_deserialization.yaml (T-X04 / CWE-502 / A08:2021): 15 sources / 12 sinks / 13 sanitizers, qualified_only: true. Closes the W372-research first-ship trio (W373 python-ssti + W374 java-sqli + W375 java-deserialization).
• W486 — Shared emit_vsa helper module. New src/roam/attest/emit_vsa.py (339 lines). cmd_pr_bundle and cmd_cga VSA emit paths collapse to 9-line + 24-line delegations. 143/143 tests pass.
• W498 — End-to-end VSA parity test. New TestVsaCliParity block in tests/test_attest_vsa.py:661+ exercising pr-bundle emit --slsa-l3 against cga emit --also-vsa for byte-identical VSA predicates. Found real drift: pr-bundle drops commit_sha when --no-auto-collect; cga path falls back to git rev-parse HEAD. Spawned W509 (fix in flight).
• W428 — Standards crosswalk YAML additions (NIST AI 600-1 + SP 800-218A). Five entries: AI600_VALUE_CHAIN_PROVENANCE, AI600_STOP_BUILD_AUTHORITY, SSDF218A_CODE_PROVENANCE, SSDF218A_CODE_REVIEW_AI_OUTPUT, SSDF218A_DEVELOPER_AUTHORIZATION. Consumes W360-research. CAISI deliberately held to H2 2026 pending substrate maturity. W506 in flight to add the missing SLSA entries (claim-integrity follow-on).
• W505-bundle — Closed-enum lints (W502 / W503 / W504). Three drift-guard lints landed together: source_framework, pass_condition, surface. 19 + 31 tests pass. Same shape as W332 / W282 / W211 vocabulary-freeze discipline.
• W482 — roam doctor ci-setup advisory check. Compared .github/workflows/roam.yml (or equivalent) against the canonical CI template. Chose advisory-check inside the existing doctor surface over a standalone roam ci-doctor command — lower friction, single dogfood touchpoint. Real-world signal: roam-code's own roam.yml has drifted from template (26 vs 28 lines) — surfaced on the dogfooded doctor run, queued for follow-on cleanup. 9 new tests + 137/137 focused pass.

Fixed — TWO long-latent silent no-ops (W493 + W499)

• W493 — taint propagation kind='calls' vs kind='call' read-side typo (CRITICAL CORRECTNESS). propagate_taint queried kind='calls' but the writers emit kind='call'. The taint DFS has been a NO-OP since inception. All 76 production findings on roam-code were stuck at chain_length=1. Three read-side sites repaired: src/roam/security/taint.py:491, src/roam/commands/cmd_dead.py:1565, src/roam/security/dataflow.py:329. Four stale tests that asserted the no-op behavior flipped to assert the real contract. Hash-stability 31/31 byte-identical. 292 tests pass. W441's 607-finding projection stands for the production roam-code corpus.
• W499 — critique/checks.py:399 impact-gate caller-edge typo (CRITICAL CLAIM-INTEGRITY). Same edge-kind typo class as W493 in a different read-site. Pre-fix matched 0 / 14,949 caller edges on roam-code — the impact gate was a COMPLETE NO-OP. Post-fix surfaces 5 high-severity findings on production roam-code. PRs touching open_db / json_envelope / to_json / invoke_cli / path now correctly exit-5 in --ci mode.

Changed — perf-measurement reframe (W485) + investigation closures (W488)

• W485 — effects_taint MEASUREMENT DRIFT (not regression). W408 baseline ran on a 17k-symbol corpus; current roam-code is 23.6k symbols / 29.9k edges / 3.8k files (+39% / +76% / 7x). Effects_taint scaled 67.6s → 87.4s; relative dominance held 48% → 50.5%. Honest reframe: the perf trajectory is intact, the apparent regression is corpus growth.
• **W488 — Sweep of remaining test_taint_*.py for stale bare-name assertions. CLEAN — W479 caught the only offender. 128 + 31 tests pass.** Closes the W479 audit's residual question.

Research / planning — W515 batch

• W441 — BAILED with a high-impact find. While investigating effects_taint Phase 2 → Phase 5 cache slice, surfaced the W493 kind='calls' vs kind='call' typo. Real wallclock when fed correct data: 0.06s — W433-research's 35s prediction was built on stale (no-op) code. Spawned the critical W493 fix. Bail was the right move (carry-over from "investigate-first bails" discipline).

> SLSA SRC-L3 evidence pipeline + Pattern-3b consolidation + taint precision discipline + perf ground-truth (W430-W491, 2026-05-15). > The wave between the W466 consolidation and this one shipped eight > threads in parallel. (1) SLSA SRC-L3 evidence pipeline end-to-end > — W451 wired the SRC-L3 lift through new src/roam/attest/vsa.py > (369 lines) and pr-bundle emit --slsa-l3 --sign --keyless; > cosign_sign_statement was already predicate-agnostic so no engine > changes were needed. W471 auto-triggered the SRC-L3 VSA emit in CI > via new template src/roam/templates/ci/slsa-src-l3.yml and the > --with-slsa-l3 flag on cmd_ci_setup, closing Gap A from W358-research. > W472 added roam cga emit --also-vsa (110-line _emit_vsa_sibling > helper) threading --sign --keyless. 23+144 / 15+31+23 / 3+43+26+43+31 > tests pass across the trio. (2) Pattern-3b consolidation closes — > W430 renamed target → symbol on 9 MCP wrappers (prepare_change, > trace, affected_tests, annotate_symbol, get_annotations, generate_plan, > get_invariants, why_fail, metrics); _PRE_W332_EXEMPT dropped 14 → 5. > Legacy target still resolves via alias with summary.alias_warnings > for back-compat. 3014 tests pass. (3) Taint engine precision > discipline reinforced — W467 fixed the W454 qualified_only bug > (root cause was a compound A+C: bare names matched via exact > qualified_name = ? on Python top-level AND via suffix LIKE '%.{name}' > on Java wrappers; fix: bare names become no-ops under qualified_only=true). > java-sqli YAML scrubbed. 125+31 tests pass. W479 audited the > remaining 22 taint YAMLs — zero offending rules — added a load-time > warnings.warn lint + 7-test hygiene guard, and drive-by-fixed an NTFS > case-collision bug (closes the open W468 + W477 items). (4) Perf > optimization ground-truth — W440 shipped the Phase 2 → Phase 5 > source-cache handoff: effects_taint moved from 91.0s → 84.7s = 7% > reduction (modest vs the 15-30s predicted by W433-research). 216 tests > pass. W441 + W485 follow-ons queued. (5) Detector FP-rate methodology > research — W470-research (dev/DETECTOR-FP-RATE-METHODOLOGY-2026-05-15.md) <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> > scoped FP-rate measurement for 3 first-to-measure detectors (smells 3047 > findings, vibe-check 831, taint). Surprise finding: OWASP Benchmark is > community-rejected — task-specific real-codebase corpora are now > preferred. Docs-only consolidation in this batch (W491); hash-stability > mandate held across all source waves.

Added — SLSA SRC-L3 wire-up + 9-wrapper rename + CI auto-trigger + cga --also-vsa (W491 batch)

• W451 — SLSA SRC-L3 wire-up. New src/roam/attest/vsa.py (369 lines) + roam pr-bundle emit --slsa-l3 --sign --keyless. cosign_sign_statement at attest/cga.py:495-594 was already predicate-agnostic, so no engine extension was required. Closes the W358-research SRC-L3 "one wave away" prediction. 23 + 144 tests pass.
• W430 — Pattern-3b 9-wrapper rename (target → symbol). Renamed across 9 MCP wrappers: prepare_change, trace, affected_tests, annotate_symbol, get_annotations, generate_plan, get_invariants, why_fail, metrics. _PRE_W332_EXEMPT dropped 14 → 5. Legacy target still resolves via _PARAM_ALIASES with summary.alias_warnings surfaced on use. 3014 tests pass.
• W471 — CI auto-trigger SLSA SRC-L3 VSA emit. New template src/roam/templates/ci/slsa-src-l3.yml + --with-slsa-l3 flag on cmd_ci_setup. Closes Gap A from W358-research (the CI-side half of the SRC-L3 evidence pipeline). 15 + 31 + 23 tests pass.
• W472 — roam cga emit --also-vsa flag. 110-line _emit_vsa_sibling helper threads --sign --keyless from the parent cga emit invocation. 3 new + 43 + 26 + 43 + 31 tests pass.
• W479 — Taint YAML qualified-name hygiene audit + lint. Audited the other 22 taint YAML rule packs and found zero offending rules. Added a load-time warnings.warn lint + 7-test hygiene guard so the regression class fails at engine load time, not on recall-limited false negatives in CI. Drive-by-fixed an NTFS case-collision bug — closes both W468 and W477.
• W440 — Phase 2 → Phase 5 source-cache handoff. effects_taint moved from 91.0s → 84.7s = 7% reduction on roam-code itself. Below the 15-30s predicted by W433-research (the savings landed on cache-warm runs rather than the headline cold path). 216 tests pass. Follow-ons W441 + W485 queued.

Fixed — W491 batch

• W467 — W454 qualified_only bug fix. Root cause was a compound A+C: bare names matched via exact qualified_name = ? (Python top-level wrappers) AND via suffix LIKE '%.{name}' (Java wrappers). Fix: bare names become no-ops when qualified_only=true. java-sqli YAML scrubbed to remove the offending bare-name entries. 125 + 31 tests pass.

Research / planning — W491 batch

• W470-research — Detector FP-rate measurement methodology. Memo at dev/DETECTOR-FP-RATE-METHODOLOGY-2026-05-15.md. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> Three first-to-measure detectors scoped: smells (3047 findings), vibe-check (831), taint. Surprise: OWASP Benchmark is community-rejected — task-specific real-codebase corpora are now the preferred evaluation substrate (Mahmoudi-class study design, not synthetic Juliet-style suites).

> Standards crosswalk research + taint rule pack v1 + shallow git default + auto-generated MCP tool table + qualified-name rule flag (W405-W466, 2026-05-15). > The wave between the W436 consolidation and this one ran twelve > threads in parallel across five families. (1) Standards > crosswalk research trilogy — W358-research (SLSA v1.2 > Source Track positioning, dev/SLSA-V12-POSITIONING-2026-05-15.md) <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> > found that roam de-facto covers SRC-L2 today, and the surprise > finding is that SRC-L3 lift is one wave — cosign_sign_statement() > at attest/cga.py:495-594 is already implemented; new wave W451 > queued. W359-research (OSCAL v1.2 Control Mapping, > dev/OSCAL-V12-CONTROL-MAPPING-2026-05-15.md) found that OSCAL v1.2 > <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> > shipped a 7th model (Control Mapping) which is the zero-prereq > first emission for per-run evidence; new waves W464/W465 queued. > W360-research (already landed in W436 batch) feeds W428. > (2) Taint rule pack v1 — W373 (python-ssti, T-X01, CWE-94; > engine already supports qualified-name matching; 7 new + 45+39 > existing tests pass) + W374 (java-sqli, CWE-89; same recall-limited > precision profile as java-fileupload because engine lacks Java > qualified-name resolution; 7 new + 44+31 existing tests pass) + > W454 (per-rule qualified_only flag for taint engine; java-sqli > opts in; 29+60 focused tests pass). Drive-bys W452-W463 queued. > (3) Perf — shallow git default on first index — W405 shipped > the 365-day shallow window via _DEFAULT_SINCE in git_stats.py + > --full-history opt-out + ROAM_GIT_SINCE env var; _first_index() > gate preserves existing deep indexes; 30+31+115 focused tests pass. > Drive-bys W437/W438/W439 queued. > (4) Documentation count drift sealed — W443 added README > coverage for 4 untracked CLI commands (evidence-diff, > evidence-doctor, llm-smells, findings); the > test_readme_covers_all_canonical_cli_commands drift guard now > passes. W449 auto-generated the README MCP tool table via a > new surface_counts.mcp_tool_descriptions() helper — 74 missing > tools added and the core preset count corrected (25 → 57). 4/4 + > 16/16 + 8/8 + 31/31 test suites pass. Drive-bys W449-W463 queued. > (5) Dedup + small-cleanup bundle — W432 removed five > oracle wrappers that W306 had already added (symbol_exists, > route_exists, is_test_only, is_reachable_from_entry, is_clone_of); > 228 → 223 decorations now match the CLAUDE.md headline. New AST > duplicate-name CI lint via surface_counts.mcp_tool_decorations() > helper. W429 packaged the W422 deprecate-permit-wrapper + > W425 lease warnings_out + W426 constitution-unparseable warning > as a single small-cleanup bundle; 204/204 tests pass; 31/31 hash > stability byte-identical. Drive-bys W443/W444/W445 and W446/W447/W448 > queued. (6) Perf research scoping — W433-research > (dev/EFFECTS-TAINT-PERF-RESEARCH-2026-05-15.md) scoped three > <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> > optimization candidates for the W408 finding: (C) double-parse I/O > elimination 15-30s zero risk; (B) function-summary memoization > 35→5s; (A) file-signature cache warm-reindex 0s. Surprise > finding: roam has TWO independent taint engines — > analysis/taint.py for Phase 5 (indexer-side) vs > security/taint_engine.py for the roam taint command — > consolidating them is a deeper structural play. Docs-only > consolidation in this batch (W466); hash-stability mandate held > across all source waves.

> Permit unification + Pattern-3b extension + llm-smells v1.1 + phase-timing reality check (W347-W436, 2026-05-15). > The wave between the W418 consolidation and this one ran nine > threads in parallel. W377-batch closed six permit-persist > red-team gaps (W377-W382) surfaced by W349; 31/31 golden hashes > remained byte-identical and 163 focused tests pass. W383 > unified pr-bundle and pr-replay permit readers behind a > single canonical roam.permits.store.load_permits_from_disk > reader, with two drive-bys captured as W421/W422. W347 extended > the Pattern-3b parameter-alias normalization to add file_path → > path (the prefix-pattern cluster was deliberately bailed on; 3 > drive-bys queued as W430/W431/W432). W415b shipped llm-smells > v1.1.0 — five new CHEAP detectors (missing_timeout, > missing_max_retries, no_system_message, no_retry_backoff, > call_in_loop); 36/36 pass; package version bumped 1.0.0 → 1.1.0; > 3 drive-bys queued as W415c/W415d/W427. W408 instrumented > per-phase timing in roam doctor and the real-data finding is the > headline of this wave: effects_taint consumes 48% of indexer > wallclock (67.6s of 139.6s on roam-code itself), which > invalidates the PageRank-first ranking in the W395-followup > perf memo; new wave W433 is queued to target effects_taint > first (drive-bys W434/W435 follow). W421 investigation bailed > after finding constitution + lease gatherers already delegate to > canonical readers (119/119 baseline tests pass; 2 drive-bys as > W425/W426). Research-only artifacts: W372-research OWASP 2026 > taint rule pack (3 first-ship rules W373/W374/W375), > W395-followup Phase 4-7 perf research (W407 reclassified to > VALIDATE — Louvain cache already implemented; top 3 new perf waves > W423/W424 + W433), and W360-research standards crosswalk > additions (5 NIST AI 600-1 + SP 800-218A YAML entries; CAISI held > until H2 2026; implementation as W428).

> MCP wrapper backfill near-complete + detector strengthening Round 2 + perf research + llm-smells design (W303-W418, 2026-05-15). > The wave between the W398 consolidation and this one moved Wave29 > wrapper backfill from 38 → 16 missing through three consecutive > sub-waves (W303 test-surface +5, W304 agent-OS daily flow +10, > W305 reports/audit +11) — 26 wrappers added in total; W306 will > drop the remaining count to ~3. Detector strengthening Round 2 > landed against the W368 BEHIND list: W370 smells empty-catch > (469 findings), W370b duplicate-conditionals (149 findings), and > W371 vibe-check modular-mirage + boilerplate-inflation > (163 + 499 findings, informational and score-preserving) — 1,280 > new findings on roam-code itself. The pitch refresh trilogy > sharpened the top-of-funnel surfaces: W390 (README + landing + > docs hero), W393 (11 secondary surfaces), W396 > (src/roam/mcp-server-card.json mirror; hash-pin updated). > Pattern-1 family Round 3 sealed cmd_owner (W362) as the third > CLI-side "exit-0 + structured envelope" fix after W327 and W324. > Permit red-team added 19 W198-edge-case tests (W349) with 6 > drive-by gaps queued as W377-W382. Structural cleanups: W346 > module-scope fixture cut test_json_contracts.py runtime ~28x; > W364 extracted _redact_secrets to src/roam/security/redact.py > (load-bearing for W363); W345 finished the W198 doc > cross-reference sweep; W319 / W348 / W352 / W403 / W412 closed > count-convention + warning-hygiene + Python-version drift + > asyncio config + stale-3.9-comment cleanup gaps; W367 refreshed > the TEAM-MCP-AUTHORITY-PRODUCT facade. Two new sonnet+web research > artifacts framed the next strategic axes: W395 perf benchmarking > (roam positioned MEDIUM — 5-20x faster than CodeQL with comparable > depth; 5 optimization sub-waves W405-W408 plus W404 scheduled) and > W402-research llm-smells pattern catalog (14 patterns; v1 = 11 > CHEAP+MODERATE — the first production-grade multi-provider > linter for openai/anthropic/google/litellm/langchain anti-patterns).

> Pitch sharpening + Pattern 1 family Round 3 + detector strengthening + permit red-teaming (W303-W398, 2026-05-15). > Four threads landed in parallel between the W375 consolidation and > the W398 one. (1) Pitch surfaces refreshed to lead with "pre-change > gates + post-change evidence": W390 swept README + landing index + > docs index hero copy, W393 extended the sweep across 11 other > surfaces (pricing / press / trust / governance / etc.). (2) > Pattern-1 family Round 3: W362 fixed cmd_owner to emit a > structured envelope on exit 0 instead of empty stdout — the third > CLI-side "exit-0 + structured envelope" fix after W327 and W324. > (3) Detector stub fills landed against the W368 BEHIND list: W370 > shipped smells empty-catch (469 findings on roam-code itself) and > W370b shipped duplicate-conditionals (149 findings; long-tail > distribution). (4) Permit red-team test surface added at W349 (19 > permit-persist tests) with 6 drive-by gaps queued as W377-W382. > Wave29 MCP wrapper backfill continued: W303 closed the test-surface > cluster (5 wrappers, 38 → 33). Structural support: W345 finished > the W198 doc cross-reference sweep, W364 extracted _redact_secrets > to a shared module (load-bearing for W363). One new sonnet+web > research artifact: W385 ecosystem positioning audit (7 tools > surveyed; 5 COMPLEMENTARY / 2 COMPETITIVE / 0 SUBSTITUTE on > agentic-assurance).

> No silent gaps milestone (W256-W261, 2026-05-14). The pr-replay > pipeline on the roam-code workspace itself now reports 7 complete > + 1 partial + 0 missing out of 8 evidence questions. Q8 > (accepted_risks / approvals) is the last open question and it is > now an explicit producer_not_available redaction-marker entry on > the packet, not a silent absence. The honest-banner classifier > (STRONG / PARTIAL / INSUFFICIENT) consumes this in the PR Replay > Markdown + JSON output, so the assurance surface can no longer > overclaim coverage it does not have.

> Pattern-1 family A/B/C/D + MCP wrapper backfill (W296-W302, 2026-05-15). > Variants A/B/C of the empty-stdout / structured-failure / hang > family are now codified as a canonical CLAUDE.md spec with 5 > invariants and external citations; Variant D (silent success on > degraded resolution) was added after W324 surfaced the gap. The > Wave29 MCP-wrapper backfill closed four clusters in four > consecutive sub-waves — exploration W299 (+9), architecture W300 > (+10), health W301 (+10), refactoring W302 (+9) — moving the > missing-wrapper count 75 → 67 → 57 → 47 → 38. Five sonnet+web > research planning artifacts landed alongside the implementation > work (Pattern-1 family audit, Pattern 3+6 audit, MCP > state-mutating patterns, standards currency audit, detector > competitive audit) — each is a forward-looking roadmap, not > shipped code.

Added

• MCP cold-start guard for index-gated tools (W296). Sealed Pattern-1 Variant A: every MCP wrapper that depends on the index now returns a structured state: "index_missing" envelope with a next_command pointing at roam init instead of hanging for ~30s on a missing .roam/. The guard runs at wrapper-entry, before any heavy import or DB connection, so cold-start latency stays below the MCP client timeout. Closes the longest-standing Variant A hang surface.
• MCP wrapper backfill — Wave29 sub-waves (W299-W302). Four consecutive sub-waves moved the missing-wrapper count 75 → 67 → 57 → 47 → 38: W299 added 9 exploration-cluster wrappers (75→67); W300 added 10 architecture-cluster wrappers (67→57); W301 added 10 health-cluster wrappers (57→47); W302 added 9 refactoring-cluster wrappers (47→38). Every new wrapper follows the W298-polish discipline (decorator audit + skip-allowlist alignment); the advisory tests/test_mcp_wrapper_coverage.py audit surfaces the remaining 38 commands plus the skip-taxonomy allowlist. Refreshed planning lives in dev/WAVE29-MCP-WRAPPER-PLAN-2026-05-15.md (W353). <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• Try-parse passthrough chokepoint sealing Variant B (W325). The MCP wrapper's _run_roam_inprocess / _run_roam_subprocess chokepoint now passes structured failure envelopes through to the LLM instead of collapsing them on a non-zero exit. Sealed Variant B for doctor, stale-refs, and test_scaffold; the pattern extends to any command that emits a structured envelope alongside an advisory non-zero exit code.
• Pattern 3 caller / complexity / rot / compliance definition fields (W331 + W331b). 9 high-signal commands now emit explicit <metric>_definition fields naming the precise computation (e.g. caller_metric_definition: "raw_edge_rows") so cross-command vocabulary drift no longer silently mismatches. W331 wired the first 6 sites; W331b closed the 3 remaining gaps and added an article-12 wording lint that blocks future drift. Vocabulary source-of-truth is the Pattern 3a/3b/6a/6b/6c codification in CLAUDE.md (W330).
• MCP input_path parameter alias normalization (W332). Four parameter aliases (file, path, paths, target_path) now canonicalise to input_path in _PARAM_ALIASES, closing the remaining cross-tool vocabulary mismatch surfaced by the Pattern-3 audit. A 1355-case AST lint pins the normalization so future wrappers cannot silently re-introduce drift.
• CALLER_METRIC_RAW canonical constant (W342). The caller_metric_definition: "raw_edge_rows" literal has been extracted to a single module-level constant and re-used across 7 emit sites (cmd_impact, cmd_preflight, cmd_understand, cmd_describe, cmd_minimap, cmd_for_refactor, cmd_invariants). W335 extended the W332 drift-guard to fail loudly on any new site that hand-rolls the literal instead of importing the constant.
• Shared _redact_secrets module (W364). Extracted from src/roam/evidence/collector.py into src/roam/security/redact.py so the redactor has a single source of truth and the evidence collector, MCP receipts, and pr-bundle emit paths all share the same regex set and allowlist. No behaviour change — the extraction was a callsite-only refactor pinned by the W232 redaction snapshot tests + a focused contract test in tests/test_security_redact.py.
• W198 permit-persist closure — doc cross-reference sweep (W345). Three HANDOVER sections, the BACKLOG Permits/Leases row, the W198 entry in section 17, and the W292/W294 historical notes were refreshed in place so the doc surface no longer describes roam permit as "verdict-facade only" once the W198 writer shipped. Pre-W198 references are preserved verbatim as historical snapshots; current-state language now reflects the writer.
• Wave29 MCP wrapper backfill — test-surface cluster (W303). Fifth consecutive Wave29 sub-wave moved the missing-wrapper count 38 → 33 by adding 5 test-surface wrappers. Trajectory across the W299-W303 arc: 75 → 67 → 57 → 47 → 38 → 33. Same W298-polish discipline (decorator audit + skip-allowlist alignment); the advisory tests/test_mcp_wrapper_coverage.py audit surfaces the remaining 33 commands plus the skip-taxonomy allowlist. W304 sub-wave is in flight against the next cluster.
• smells empty-catch detector — first stub filled (W370). The first of the W368 BEHIND-list smells stubs has been promoted from placeholder to a real detector — 469 findings emitted on the roam-code workspace itself. Detection follows the smells-helper template so the finding rows go through the canonical _emit_smells_findings path and inherit the registry's tier-mapping + version-stamp discipline. Closes the first item of the W368 audit's "smells rule depth (empty-catch + primitive-obsession are placeholder stubs)" gap.
• smells duplicate-conditionals detector — second stub filled (W370b). The second of the W368 BEHIND-list smells stubs has been promoted to a real detector — 149 findings on roam-code with a long-tail distribution (a small set of expressions account for most of the duplicate-conditional rows). Same emit path as W370; W370c will close the remaining stubs in the W368 list.
• Permit red-team test surface (W349). 19 permit-persist tests added that exercise the W198 writer's edge cases (corrupt JSON document on disk, partial write, racing writer, schema drift, expired-permit reads, missing parent directory, etc.). Six drive-by gaps surfaced and queued as W377-W382 for the next session — none block the W198 happy path, but the red-team surface is what gives the permit substrate the same producer-grade hardening as the lease and run-ledger substrates.
• cmd_owner Pattern-1 envelope fix (W362). Third CLI-side "exit 0 + structured envelope" Pattern-1 fix after W327 (pytest-fixtures) and W324 (roam_annotate_symbol). Pre-W362 roam owner <symbol> on a symbol with no owner data exited 0 with empty stdout, so the MCP wrapper crashed in json.loads(""). Post-W362 the command always emits a structured envelope; on no-owner it returns state: "no_owner_data" + next_command: "roam blame". Pattern-1 Variant C contract preserved.
• Pitch refresh — README + landing index + docs index (W390). Hero copy on the three top-of-funnel surfaces now leads with "pre-change gates + post-change evidence" — the dual framing that surfaces both halves of the agentic-assurance thesis in one line. Sweep also touched the "what roam does" callout blocks and the roam evidence doctor mention. No structural HTML change; only the lede text.
• Pitch refresh — 11 secondary surfaces (W393). Extended W390 across pricing.html, press.html, trust.html, governance.html, the four services-reports/ deliverables, and three audit-report/ templates. Same "gates + evidence" framing; same no-structural-HTML-change discipline.
• Wave29 MCP wrapper backfill — agent-OS daily flow cluster (W304). Sixth consecutive Wave29 sub-wave moved the missing-wrapper count 33 → 23 by adding 10 wrappers around the run-ledger / mode / lease / permit / memory / brief / next / agent-score daily-flow surface. Same W298-polish discipline (decorator audit + skip-allowlist alignment); the advisory tests/test_mcp_wrapper_coverage.py audit surfaces the remaining 23 commands plus the skip-taxonomy allowlist.
• Wave29 MCP wrapper backfill — reports / audit cluster (W305). Seventh consecutive Wave29 sub-wave moved the missing-wrapper count 23 → 16 by adding 11 wrappers (some commands surfaced via the same wrapper) around the reports / audit-trail / pr-bundle / evidence-doctor / replay output surface. Wave29 trajectory across the full arc W299-W305: 75 → 67 → 57 → 47 → 38 → 33 → 23 → 16. Three to four wrappers remain (W306 lands them).
• vibe-check modular-mirage + boilerplate-inflation detectors (W371). Two informational AI-rot patterns added to vibe-check. modular-mirage (163 findings on roam-code itself) flags fragmented one-line wrappers around trivial logic — a common AI shape. boilerplate-inflation (499 findings) flags excessive scaffold-style commentary and stub structure. Both are score-preserving: emitted into the findings registry with tier heuristic and do not move the health score, so adoption is zero-risk. Adds two new vibe-check kinds to the canonical 8-pattern catalog without disturbing existing ones.
• test_json_contracts.py module-scope fixture (~28x speedup) (W346). Per-test indexing was rebuilding the full DB on every contract case. Promoted to a module-scope fixture so the index is built once and reused across all contract assertions. Runtime drops from ~6 minutes to ~13 seconds on this file; broader test sweep benefits proportionally because the contracts file was the slowest single module.
• TEAM-MCP-AUTHORITY-PRODUCT facade refresh (W367). Updated dev/TEAM-MCP-AUTHORITY-PRODUCT-2026-05-14.md to reflect the <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> post-W198 reality (real permits exist; the W292/W294 corroboration harvester reads them; the local single-agent receipt model is shipping ahead of any networked Team MCP work, as the threat-model memo W214 prescribes). Documentation refresh only.
• src/roam/mcp-server-card.json pitch-refresh mirror + hash-pin update (W396). The MCP discovery card now mirrors the W390+W393 refreshed pitch ("pre-change gates + post-change evidence"). The hash-pin in tests/test_mcp_server_card_hash.py was updated in step so the canonical card stays drift-locked.
• roam findings CLI (list / show / count) — cross-detector finding registry surfaced from the new findings table in src/roam/db/findings.py. After the W95 clones migration this is mcp_expose=True + maturity stable.
• roam taint now actually fires findings on PHP/Laravel codebases — the engine BFS query in src/roam/security/taint_engine.py:283,333 was filtering on kind IN ('calls','references') but the index stores singular 'call'/'reference' edge kinds, so the entire roam taint subsystem had been silently returning 0 findings since v12. Fix is one line; affects every prior-shipped taint rule.
• PHP/Laravel taint rule pack: 5 YAML files / 101 rule entries under src/roam/security/taint_rules/php_*.yaml (command injection, Laravel open-redirect, Laravel SQLi, Laravel XSS, path traversal).
• Detector registry: @detector(...) decorator + roam math --list-detectors, --only, --exclude flags.
• Plugin contract: register_framework_detector(detect_fn: Callable[[Path], Optional[str]]) now carries the typed signature.
• mcp_introspection_available: bool field in roam surface --json envelope — consumers can now distinguish "no fastmcp installed" from "fastmcp broken".
• roam doctor advisory: "Index step missing because step X failed; run roam index --force" — reads the W82 step-completion manifest.
• Per-component VERSION stamps (bridges, detectors, extractors) — populated in edges / symbols rows + the indexer manifest.
• New landing pages: /governance (409 lines), /trust (434 lines, honest SOC 2 Q1 2027 / ISO 42001 Q3 2027 stance), pricing.html FAQ block (7 items), templates/audit-report/ (template + renderer + 3 sample reports), templates/services-reports/ (4 service deliverables: AI adoption readiness, due diligence, post-incident replay, security reachability triage).
• Findings registry — 14 detectors persisting findings via --persist (W109-W136). Migrations across the wave: smells (3047 findings on roam-code), n1, missing-index, over-fetch, bus-factor (65), auth-gaps, vulns, invariants/laws (9), hotspots, taint, vibe-check (831 — 8 AI-rot patterns), orphan-imports (344), conventions (39), pr-risk, duplicates (853). All detectors emit through the canonical _emit_<X>_findings(conn, data, source_version) template so the registry's tier-mapping and version-stamp discipline is uniform across the wave.
• Plugin substrate: register_framework_profile (W123 / Wave28.3). New FrameworkProfile dataclass bundles a framework's detect_fn + file_patterns + recommended_commands + conventions so a plugin can declare its profile in a single call. Surfaced as a method on RoamPluginContext; the reference example plugin in dev/example-plugin/ was migrated to demonstrate the contract.
• Findings-registry subject-kind vocabulary expanded. The canonical enumeration now accepts module, directory, endpoint, package, and commit in addition to the originals symbol and file. New detectors in this wave (orphan-imports, duplicates, conventions, bus-factor) consume the expanded vocabulary.
• mcp_tool_count_by_preset field on roam surface --json envelope (W138). Per-preset MCP tool counts: core 57 / review 70 / refactor 70 / debug 69 / architecture 71 / compliance 13 / full 149. Machine consumers no longer have to introspect _PRESETS to know the shape of each preset.
• OneDrive / Dropbox / iCloud cloud-sync detection at roam init (W127). Warns when .roam/ would land on a cloud-synced path (corrupts SQLite WAL files on multi-device sync). roam doctor advisory was already in place; the init-time warning catches the issue before the first index instead of after the first crash.
• Agentic-assurance pipeline — actor / authority / approvals first-class on ChangeEvidence (W189-W211). Six-step arc that turns the W174 evidence dataclasses into a portable assurance packet: pr-bundle producer emits an actor block (agent_id / human_actor / mcp_client_id / tool_id / ci_runner_id / actor_kind) plus empty approvals[] / accepted_risks[] arrays (W189); the mega collector materialises actor_refs[] / authority_refs[] / environment_refs[] from envelopes, the run ledger, and CI env (6-provider CI detection; W190); PR Replay renders Actors / Authorities / Environment sections with the assurance-leads / findings-follows ordering (W191); a vocabulary-drift sweep aliased author→actor on pr-risk + bus-factor envelopes and documented the permit-facade contract (W198); ActorRef gained a trust_tier (verified_ci / git_author / local_env / self_reported_agent / unknown), AuthorityRef gained a source enum (mode / permit / rule_config / ci_policy / human_approval / inferred_fallback) with a facade auto-stamp, ApprovalRecord graduated to a first-class dataclass with expiry, and every model carries an explicit NON-GOALS docstring (W211).
• ChangeEvidence schema extensions — 9 new optional fields (W210). Time-aware (context_read_at, edits_started_at, edits_completed_at); stale-evidence (evidence_stale: bool, stale_reasons[]); version-linked (roam_version, rules_config_hash, constitution_hash, control_map_hash). Adds two computed methods on the dataclass: assurance_floor() (lowest trust_tier across actors/authorities) and evidence_completeness() (fraction of optional evidence slots populated). Backward-compat preserved: schema_version stays "1.0.0" via _W210_OMIT_WHEN_DEFAULT_FIELDS so packets with no new fields serialise byte-identical to W174.
• MCP decision-receipt emitter (W183 / W196). Per sensitive @_tool invocation, the wrapper writes .roam/mcp_receipts/<run_id>/<tool_call>.json capturing the agent identity, the call args (after redaction), and the verdict. New McpDecisionReceipt dataclass under src/roam/evidence/ (W183); emitter wired into the decorator with best-effort discipline — a receipt write failure NEVER breaks the underlying tool call (W196).
• Mega collector extension — 5 new kwargs (W199). The W175 collector now accepts rules_envelopes, audit_trail_envelope, vuln_reach_envelopes, test_impact_envelopes, cga_envelopes, mcp_receipts_dir; each flattens into the relevant ChangeEvidence fields. Replaces the W176 audit-trail synthetic-finding stop-gap with first-class manifest artifact + policy_decisions entries.
• roam pr-replay producer wiring — 6 gatherers (W223). Wired rules / audit-trail / vuln-reach / test-impact / cga / mcp-receipts into the PR Replay producer; coverage on the roam-code workspace lifted from 3/8 to 6/8 evidence slots. The remaining two gaps (W219 producer-side) are documented in the next-session queue.
• Control mapping v1 schema (W184). YAML control-map entries now carry source_framework / evidence_types[] / surface / wording_guard fields. The CI lint at W203 caught 3 W184 drift entries (wording that overclaimed certification) and fixed them in place.
• roam evidence-diff CLI command (W225). Diffs two ChangeEvidence JSON files structurally: changed_subjects / findings / policy_decisions / actors / authorities. Used by W232's redaction snapshot tests.
• Export profiles for evidence packets (W226). Four profiles — internal / customer / audit / public — drive per-audience redaction policy. Each profile is a deterministic field-allowlist applied at projection time, so the same source packet renders 4 different views with one redaction surface.
• False-positive feedback loop module (W228). Routes user-flagged false positives back through the findings registry with provenance preserved (which detector / which version emitted the row).
• PR Replay "Evidence Limitations" section (W185). Every PR Replay Markdown report now lists which evidence slots are populated and which are absent, plus the reason for absence. Reads from the same ChangeEvidence packet that drives the body.
• Strategic memos shipped overnight. Five new dev docs landed alongside the agentic-assurance pipeline: dev/MILESTONE-2026-05-13-AGENTIC-ASSURANCE.md (W202 milestone <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> integration note); dev/THREAT-MODEL-MCP-2026-05-14.md (W214 — 6 <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> threats, conditional P-tier escalation contingent on Team MCP Gateway shipping); dev/POSITIONING-TEAM-MCP-2026-05-14.md (W215 <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> — authority-product framing); dev/DEMO-NARRATIVE-CANONICAL.md <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> (W216 — 8/8 ideal-case fixture, deterministic content_hash 17958f73…); dev/PIPELINE-REVALIDATION-2026-05-14.md (W230 — <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> end-to-end re-validation report).
• pr-replay context_refs producer (W246). New _gather_context_files in src/roam/commands/cmd_pr_replay.py:895 runs git diff --name-only ONCE for the whole commit window (not per-commit), caps emission at 500 entries with a truncation warning, and stamps pr_bundle_envelope["context_files"]. The W199 collector then materialises context_refs[] via _build_context_refs_from_context_files. Smoke run on roam-code: 492 context_refs populated. Closes evidence question Q3 (context_read) on the real-workspace pipeline. The executable 8-question audit threshold ratcheted 5 → 6 in tests/test_eight_questions_audit.py:344; 4 new tests in tests/test_evidence_pr_replay.py.
• Pipeline re-validation v3 (W254). Real-world roam pr-replay HEAD~5..HEAD on the roam-code workspace itself now scores 7 complete / 0 partial / 1 missing out of 8 evidence questions (trajectory W201 → W230 → W254: 3 → 3 → 7). Q8 (accepted_risks / approvals) remains the only producer-side gap. Synthesised collector ceiling stays at 8/8 — confirms the remaining gap is producer-side, not collector-side. Honest-banner thresholds proposed (consumed by W259): complete ≥ 7 STRONG; complete + partial ≥ 5 ∧ missing ≤ 3 PARTIAL; otherwise INSUFFICIENT.
• Producer coverage matrix (W252) — dev/PRODUCER-COVERAGE-MATRIX-2026-05-14.md (168 lines). 15 <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> tier-1 producers analysed per-field; 22 tier-2 detectors collapsed. Top three under-served fields: environment (1 producer: pr-replay only), policy (2 producers: rules + audit-trail-verify), authority (1 producer: pr-bundle.mode; permits/leases are verdict-facade only). Memo also surfaces the pr-replay-synthesises-its-own-pr_bundle gap — meaning W240's actor + scrub fixes do NOT propagate through the replay path (queued as W260).
• Executable 8-question audit threshold ratcheted 6 → 7 (W258). Synth-fixture enrichment in tests/test_eight_questions_audit.py extracted a _reconstruct_artifacts(rows) helper, reused it for both context_refs[] and artifacts[], and lifted EXPECTED_COMPLETE_COUNT_TODAY 6 → 7. Trajectory across the W220 / W246 / W258 arc: 3 → 5 → 6 → 7. The test now guards the W254 real-workspace ceiling against regression.
• Honest evidence-coverage banner (W259). New module src/roam/evidence/banner.py exports classify_evidence_coverage() / render_banner_markdown() / banner_envelope_block() driving the three-tier classification (STRONG complete ≥ 7; PARTIAL complete + partial ≥ 5 ∧ missing ≤ 3; INSUFFICIENT otherwise). Wired into PR Replay Markdown at cmd_pr_replay.py:1444 (above the W185 limitations section) and into the JSON envelope as extra_payload.evidence_coverage + the summary.evidence_coverage_tier mirror. Template placeholder {{evidence_coverage_banner}} at templates/audit-report/pr-replay-template.md:3. 5 new banner tests; 90/90 focused tests pass. Closes the W254 / section 11.4 banner proposal.
• actor_helpers.py extraction + pr-replay synth-bundle parity (W260). New src/roam/commands/actor_helpers.py exports resolve_actor_block() + resolve_actor_kind() with the W189-canonical resolution priority (CLI flag > env var > git config > active run-ledger agent). cmd_pr_bundle.py's symbols are now thin back-compat wrappers (37/37 tests/test_pr_bundle.py preserved). cmd_pr_replay.py calls the shared helpers and re-runs W249's _scrub_actor_block immediately after resolution on the producer side, so the synth pr_bundle envelope is now byte-equivalent to a real cmd_pr_bundle emission for actor + scrub purposes. 3 new tests; 86/86 focused tests pass. Closes the W252-surfaced pr-replay-bypasses-cmd_pr_bundle gap (HANDOVER section 11.3).
• environment_refs shared helper + pr-bundle wiring (W266). New module src/roam/evidence/env_refs.py exports build_environment_refs(*, commit_range=None, workspace_root=None, env=None) -> tuple[EnvironmentRef, ...], giving producers a public API for assembling the environment axis in canonical order (ci_job → workspace → branch_range → local_run). Strategy: delegate, not move — the collector's pre-existing _build_environment_refs has 30+ call sites and the v0/v1 content-hash contract is tested against its exact output, so the helper delegates CI detection to the collector's _detect_ci_env_id (the W251 6-provider env-var matrix stays the single source of truth) and implements its own producer-friendly assembly path on top. cmd_pr_bundle.py's _build_envelope now stamps environment_refs[] on every emit path (init / set / add / emit / validate); environment_refs added to the _PR_BUNDLE_KNOWN_PAYLOAD allowlist. 11 helper tests + 2 pr-bundle tests; 101 focused + 77 evidence/collector + 32 pr-bundle pass. Smoke on roam-code emits real env_refs (workspace path + commit + hostname). Closes the W252 environment-axis gap — 1 producer → N.
• pr-bundle permits/leases real producer (W268). Two new helpers in cmd_pr_bundle.py:1444-1532: _load_permits_from_disk(repo_root) scans .roam/permits/*.json and returns [] when the directory is absent (at the time W268 shipped, the historical permit facade was not yet persisted — Pattern-2 always-emit contract preserved; W198 has since shipped the writer side, so this path now returns real rows on workspaces that have issued permits via roam permit issue --persist); _load_leases_from_disk(repo_root) delegates to roam.leases.list_leases(include_expired=True, include_released=True) so the on-disk schema stays single-sourced through Lease.to_dict(). Both invocations reuse the find_project_root() result already computed for environment_refs, and the results stamp top-level permits=permits_out, leases=leases_out on the envelope. The collector's pre-existing _build_authority_refs (lines 946-954) now sees real permit + lease rows and produces AuthorityRef(authority_kind="permit", …) / …="lease", … entries: pre-W268 the authority axis carried 1 ref kind (mode); post-W268 it carries 5 (mode + permit + lease + policy_rule + approval), with permit + lease backed by real producer rows. Hash-stability proven — 31/31 golden hashes pass (envelope-level fields flow into already-omit-when-empty authority_refs, so ChangeEvidence.content_hash is unaffected). 6 new tests (4 pr-bundle + 2 collector); 160 focused + 56 broader pass. Smoke on roam-code: permits=[] (no .roam/permits/ yet), leases= 2 entries with full schema. Closes the W252 authority-axis gap — 1 producer → 5 ref kinds.

> Three closures in three waves milestone (W261 + W266 + W268, > 2026-05-14). Three of the W252 producer-coverage matrix's > top-three under-served fields closed in three consecutive waves: > Q8 silent-gap → explicit producer_not_available redaction > marker (W261); environment axis 1 producer → N (W266); > authority axis 1 producer → 5 ref kinds (W268). All three > followed the same discipline — Pattern-2 always-emit + > delegate-not-move — so existing v0/v1 content-hash contracts > survived intact, and consumers see the new evidence the moment > a producer ships without a schema migration or flag flip.

• pr-replay policy adapters for constitution / permits / leases (W267). Three new gatherers in cmd_pr_replay.py: _gather_constitution_policy_decisions (~line 995), _gather_permit_policy_decisions (~line 1056), _gather_lease_policy_decisions (~line 1117). Wiring strategy (option b): the collector's collect_change_evidence gained an extra_policy_decisions kwarg so additional rule sources project into the canonical packet rather than bypassing it (canonical mandate). Concatenation order is stable — rules → audit-trail → extras — so content_hash stability is proven (31/31 golden hashes pass). Smoke on roam-code: 1 audit-trail + 3 constitution gates + 2 leases + 0 permits = 6 policy_decisions on the canonical packet, up from 1 pre-W267. 5 new tests; 127/127 focused + 71/71 broader pass. Closes the W252 policy-axis gap — 2 producers → 6 decisions from 4 sources.
• pr-replay synth-bundle full parity for permits / leases / env_refs (W272). Strategy A — direct import, no new module — proved sufficient: _load_permits_from_disk / _load_leases_from_disk were already clean module-level functions in cmd_pr_bundle.py and only needed a docstring annotation to flag them as shared. The W272 stamping block lives at cmd_pr_replay.py:1378-1466 immediately after the W260 actor block; the post-collector environment_refs merge sits at :1577-1601. Dedup decision (option c): stamp permits / leases on the synth pr_bundle envelope for direct-consumer parity, and merge the W266-built env_refs tuple onto packet.environment_refs after the collector returns (the collector ignores the envelope's environment_refs key when rebuilding from raw inputs — naive stamping would have lost the workspace ref). Merge dedupes by (env_kind, env_id). Smoke on roam-code: 3 actor_refs (W260), 3 authority_refs (mode + 2 leases — up from 1), 3 environment_refs (branch_range + local_run + workspace — up from 1). 4 new tests; 146/146 focused + 72/72 broader pass. Closes the W260 + W266 + W268 producer-parity gap on the replay path.

> W252 producer coverage matrix closure cycle complete > (2026-05-14). Four waves collapsed the matrix's top-three > under-served fields in a single cycle: environment (W266), > authority (W268), policy (W267), and synth-bundle parity (W272). > Real-world roam pr-replay HEAD~5..HEAD on the roam-code > workspace itself now carries: 3 actor_refs / 3 authority_refs / > 3 environment_refs / 6 policy_decisions / 492 context_refs / > 11 artifacts (audit-trail manifest + 10 CGA predicates); the > executable 8-question audit reports complete=7 / partial=1 / > missing=0 (W261 forward-compatible no-silent-gaps shape). The > last open producer-side gap (Q8 accepted_risks / approvals) > is the only path to lift the partial → complete and is queued > P1 as W247.

• roam permit issue --persist writer — permit-persist closure (W198). Closes the W186 audit's verdict-facade gap on roam permit. Pre-W198 the command was strictly a verdict facade: no document was written to .roam/permits/<permit_id>.json, so the W268 collector's _load_permits_from_disk always returned [] (Pattern-2 always-emit contract held). W198 ships the writer side via a new roam permit issue subcommand on top of the existing verdict-facade command + a new disk-backed src/roam/permits/store.py store: with --persist, writes one JSON document per issued permit to .roam/permits/<permit_id>.json; without --persist, the command remains a dry-run and stays byte-stable with the pre-W198 verdict contract (hook / pre-commit gates that consume only the verdict are unaffected). Closes the W292 / W294 facade-vs-real corroboration loop: the W292 harvester now finds real permit rows on disk, and AuthorityRef(authority_kind="permit", …) entries carry a real extra["permit_id"] with provenance="producer_envelope(permit)" instead of the historical inferred-fallback marker. The W294 authority-source mapping (permit → "permit") now resolves on real permit rows rather than synthetic envelope fields. 95 permit-persist-focused tests in tests/test_cmd_permit_persist.py + 105 broader pass; existing W292 / W294 authority-provenance goldens stay byte-identical (no schema change — the permit row is an additional populated path, not a new field default).

• Packet size budget enforcement on ChangeEvidence (W280). New module-level constant PACKET_SIZE_BUDGET_BYTES = 262144 (256 KiB) measured on canonical JSON. _apply_size_budget() is called BEFORE with_content_hash() so the content_hash is the hash of the post-truncation packet — never of a packet that the consumer will not see. Frozen deterministic 5-step truncation order: artifacts.content_inline → context_refs.content_inline → policy_decisions.extra → findings.evidence → actor_refs.extra. Redactions are NEVER dropped (they describe the truncation itself). The "size_limit" redaction reason is appended dedup-safe. roam evidence doctor surfaces a packet_size: {bytes, budget_bytes, budget_state} block, with oversized_after_truncation mapping to WARN (not FAIL). First enforced size discipline anywhere in the evidence pipeline; real packets on roam-code sit at ~96 KB (~37% of budget).
• Provenance vocabulary (vocab + helper; wiring deferred) (W282). New PROVENANCE_SOURCES frozenset (10 values: ci_env_var / git_config / run_ledger / cli_flag / env_var / producer_envelope / audit_trail / mcp_receipt / inferred / unknown) plus provenance_label() pure helper with a detail-compact form. Cross-vocabulary leakage validation pins the new enum against the existing assurance vocabularies. CLAUDE.md vocab table grew 11 → 12 rows. Producer-side wiring is deliberately deferred to a future wave (W290+) so the vocabulary lands clean before any call-site churn.
• Limitations generated from packet structure on pr-replay (W284). _derive_limitations(evidence) at cmd_pr_replay.py:2295 projects three packet sources into the Markdown / JSON evidence_limitations block, in a frozen deterministic order: Q-gaps (Q1 → Q8) → redactions (tuple order) → trust-tier warnings (actor_refs order) → non-cert footer always appended. New _Q_GAP_LABELS + _REDACTION_EXPLANATIONS lookup tables. Renderer _render_evidence_limitations() at :2136 rewrites the prior hand-written boilerplate into a pure structural projection. Sentinel _No evidence limitations detected._ when no source contributes. Closes the "limitations drift from packet" hazard that surfaced during the W281 trust-tier rollout.
• roam_version stamped at the producer site on ChangeEvidence (W287). New _resolve_roam_version() helper at change_evidence.py:783 (deferred import of roam.__version__, fallback "unknown"). Wired at the producer site in collect_change_evidence() at collector.py:2667 — deliberately NOT as a dataclass field default, which would have broken the W210 omit-when-None invariant for packets that do not opt into the field. Backward compatibility: pre-W210 packets and W210 packets that do not populate the new fields still serialize to byte-identical canonical JSON; existing stored content_hash values stay valid. Smoke on roam-code stamps the live PyPI version ("13.0").
• GitHub PR review parser / normalizer (W247a). New module src/roam/evidence/github_reviews.py (~360 lines) introduces three public functions: parse_github_reviews() (pure; deterministic input → output), load_reviews_from_fixture() (pure offline path for tests), and harvest_reviews_from_gh_cli() (deliberate opt-in subprocess path). New GITHUB_REVIEW_STATES closed enum (5 values). Discipline: APPROVED reviews land in approvals[] only when the review's commit_id matches head_commit_sha (stale approvals are filtered); CHANGES_REQUESTED reviews land as PolicyDecision(decision="deny") rows so the deny signal flows into policy_decisions[]; COMMENTED / DISMISSED / PENDING reviews are filtered with explicit warnings. Review bodies are NEVER stored (asserted by test). Fixture-first design keeps the test suite fully offline. This is the first half of the W247 real approvals producer — W247b (pr-replay integration) is queued separately so the parser proves itself before any consumer wiring.
• policy_decisions and approvals provenance stamping at 9 ingestion points + Pattern-2 fallback (W293). Every PolicyDecision row + every ApprovalRecord row now carries extra["provenance"] = provenance_label(source, detail=...) built from the W282 closed PROVENANCE_SOURCES vocabulary (no new strings invented). Producer-side stamping fires at the gatherer / parser / CLI command — never at the dataclass default — so existing producer-supplied provenance is preserved idempotently. Ingestion sites: cmd_pr_replay.py:1061-1086 constitution gatherer → producer_envelope(constitution); cmd_pr_replay.py:1120-1153 permit gatherer → producer_envelope(permit); cmd_pr_replay.py:1198-1224 lease gatherer → producer_envelope(lease); cmd_pr_replay.py:1267-1280 approval flattener → producer_envelope(github_review); github_reviews.py:370-380 PolicyDecision builder → producer_envelope(github_review); collector.py:2459-2467,2497,2502 audit-trail chain-integrity → audit_trail; collector.py:1973-1976,2018 rules envelope → producer_envelope(rule); cmd_pr_bundle.py:2399-2410 add-approval CLI → cli_flag; collector.py:3361-3389 Pattern-2 fallback stamps unknown only when no upstream signal exists (existing values preserved). Sharp drive-by: PolicyDecision.to_dict() flattens extra to a top-level key on the wire and from_dict() re-nests it on read, so wire format stays flat while the in-memory dataclass stays typed — ChangeEvidence.__post_init__ handles both shapes. The W247a body-prohibition guardrail (test_review_bodies_do_not_appear_in_canonical_json) extended to assert no body keys appear on any row after the provenance hop. Smoke on roam-code: 6 policy_decisions all carry explicit provenance (3 constitution + 2 lease + 1 audit_trail; zero unknown fired). 15 new tests + 1 regression assertion + 486 broader pass + 31/31 goldens byte-identical.
• AuthorityRef.source population + auto_log() writer-side run-ledger event fields (W294). Closes both W292 follow-ups in one wave. (a) AuthorityRef.source now populated DISTINCTLY per category mapping at _build_authority_refs via a new source= kwarg on the _add helper: mode → "mode"; permit → "permit" + extra["permit_id"] when a real permit row is present; policy_rule → "rule_config"; approval → "human_approval"; lease intentionally retains "inferred_fallback" because AUTHORITY_SOURCES has no lease literal (deliberate vocab decision, documented inline — see HANDOVER §16.5 lease asymmetry note). (b) auto_log() in src/roam/runs/helpers.py gained an optional extra_event_fields kwarg with a closed whitelist _AUTHORITY_EVENT_FIELDS = {"mode", "active_mode", "mode_to", "mode_from", "permit_id", "lease_id", "approval_id", "rule_id"}. Writer-side wiring: cmd_mode (:362-386) emits mode_to + mode_from on non-noop switch (with pre- switch capture before set_active_mode runs); cmd_lease (:329-340 claim, :430-441 release) emits lease_id; cmd_pr_bundle add-approval (:2392-2403) emits approval_id when an active run exists. AuthorityRef.source (W211 category) and extra["provenance"] (W282 channel) remain DISTINCT load-bearing fields — neither is a synonym of the other. 15 new tests + 1 updated assertion (W292's source assertion flipped from "inferred_fallback" to "mode" now that W294 distinguishes) + 31/31 goldens byte-identical + 305 broader pass.
• EvidenceArtifact advisory warning when content_inline > 8 KiB (W288-followup). Construction-time warnings.warn at EvidenceArtifact.__post_init__ fires whenever len(content_inline) > INLINE_CONTENT_SOFT_LIMIT_BYTES (8 KiB). The limit stays purely advisory — no reject, no truncate, no redaction stamp; consumers can still build large inline artifacts when intentional. Companion to W280's enforced 256 KiB packet-level budget: a two-tier discipline where the per-artifact limit is a pressure signal and the packet-level limit is the hard ceiling that drives the deterministic truncation order. Focused tests green.

> Provenance trilogy closure (W290 + W292 + W293, 2026-05-15). > Every evidence dimension (actor_refs, authority_refs, > policy_decisions, approvals) now carries > extra["provenance"] stamped at ingestion sites via > provenance_label() from the W282 closed vocabulary (10 > sources, no new strings). Ingestion-point stamping discipline > preserves dataclass schema cleanliness — every wave landed > through call-site additions, never through default-value > changes. W294 stabilized the authority axis by populating > AuthorityRef.source distinctly per category and wiring > writer-side run-ledger fields (mode_to/mode_from/ > permit_id/lease_id/approval_id/rule_id) so the W292 > harvester finds real corroboration instead of only > run-meta.mode. W288-followup added a per-artifact advisory > warning to complement W280's enforced packet-level budget. 31/31 > golden content_hashes byte-identical across the trilogy + > stabilization waves.

Changed

• cmd_owner Pattern-1 envelope contract (W362). Behaviour change at the JSON envelope boundary: roam owner <symbol> on a symbol with no owner data previously exited 0 with empty stdout (Pattern-1 Variant C crash surface). It now emits a structured envelope with state: "no_owner_data" + next_command: "roam blame" while still exiting 0. The MCP wrapper round-trip is now crash-free; downstream consumers reading state and next_command pick up the new contract automatically.
• Python 3.10+ minimum documented across the source tree (W352). Pyproject already required 3.10+ but a long tail of source-file comments and contributor docs still referenced 3.9 compatibility. Sweep aligned the comment/doc surface with the actual requirement — no functional change. Companion to W412 which removed the last stale from __future__ import annotations rationale notes that predated the 3.10+ baseline.
• Closed plugin_count convention drift (W319). The plugin count (currently 1 — the reference example plugin) is now sourced from a single canonical citation. Closes a Pattern-3-style drift where the count was written in 3 places with different values.
• W288 INLINE_CONTENT_SOFT_LIMIT advisory wording polish (W348). The advisory warnings.warn text was tightened so consumers don't conflate the soft 8 KiB advisory with the hard 256 KiB packet-level budget. Pure wording polish; the threshold logic and the W288-followup contract are unchanged.
• asyncio configuration cleanup (W403). Pyproject's pytest config no longer carries a stale asyncio_mode entry that predated a dependency drop. Test runtime is unchanged; the warning emitted on every test run is gone.
• CLAUDE.md Pattern-1 family expanded to A/B/C (W328). The empty-stdout / structured-failure / hang family is now codified as a canonical three-variant spec: A (hang on missing prerequisite), B (structured signal lost via exit-code gap), C (empty stdout crashes json.loads). Five invariants pin the contract (always-emit envelope, structured is_error, cold-start guard, exit-code passthrough, json-parse-on-empty defence) and external citations point at the FastMCP / Anthropic / MCP-spec sources that drove each invariant.
• CLAUDE.md Pattern 3 + Pattern 6 expanded (W330). Pattern 3 now carries explicit 3a (cross-command metric drift) + 3b (cross-tool parameter-name drift) variants; Pattern 6 grew to 6a (response volume crisis), 6b (token-count drift across recipes), and 6c (handle-pattern crash on its own handles). External citations point at agi-in-md LAW-4 evidence and at the in-tree W332 / W325 / W331 implementation waves so future maintainers can trace the codification back to the experiments that produced it.
• Pattern-1 family Variant D added (W334). After W324 surfaced a silent-success failure in roam_annotate_symbol (the wrapper returned verdict: "completed" despite the underlying resolver degrading to a no-match), Variant D was added to the canonical spec: "silent success on degraded resolution." The variant is distinct from Variant B (structured signal lost via exit-code gap) because Variant D's failure mode is producer-side — the command itself fails to surface the degradation, not the MCP wrapper. The fix lives at the command's resolver boundary, not the wrapper chokepoint.
• authority_refs provenance stamping with deterministic precedence ladder (W292). Mirrors the W290 actor_refs pattern: each AuthorityRef now carries extra["provenance"] = provenance_label(source, detail=...) built from PROVENANCE_SOURCES. New _resolve_authority_provenance() at src/roam/evidence/collector.py:1063 + _collect_corroborated_authorities_from_runs at :967 + rewritten _build_authority_refs at :1153. Frozen precedence ladder on same (authority_kind, authority_id): run_ledger > audit_trail > mcp_receipt > producer_envelope(permit) > producer_envelope(mode) > producer_envelope(rule) > producer_envelope(approval) > producer_envelope(lease) > generic producer_envelope > inferred > unknown. Implemented as a deterministic ladder in _resolve_authority_provenance, NOT as implicit dict iteration — HMAC-verified run-ledger evidence always beats any envelope claim. AuthorityRef.source (W211 AUTHORITY_SOURCES — the category answer) and AuthorityRef.extra["provenance"] (W282 PROVENANCE_SOURCES — the channel answer) preserved as INDEPENDENTLY load-bearing fields; merging them would erase information. 14 new tests + 31/31 goldens byte-identical + 180 broader pass.
• DB schema USER_VERSION 13 → 17 (four migrations during the extension: W82 step manifest → 14, W81 version stamps → 15, W89 findings table → 16, W97 FTS5 schema-hash discipline → 17).
• _DESTRUCTIVE_TOOLS in src/roam/mcp_server.py:301-307 is now a derived view (frozenset[str]) built from the @_tool(destructive=True) decorator kwarg — one source of truth instead of a hand-maintained dict.
• cmd_findings maturity flipped experimental → stable + mcp_expose=True after W95 (clones detector migration emitted 584 findings on roam-code).
• cmd_surface.py mcp_tool_count source-of-truth flipped to _TOOL_METADATA (W138). Previously read _REGISTERED_TOOLS, which is empty when fastmcp isn't installed — so surface --json reported mcp_tool_count: 0 on stripped installs. _TOOL_METADATA is env-independent (it's built at import time from @_tool decorator kwargs). mcp_introspection_available remains as a separate signal so consumers can still distinguish "stripped install" from "broken fastmcp".
• A1 dict-collapse ladder progress: 6 of 8 dicts now derived. _NON_READ_ONLY_TOOLS (W108) and _NON_IDEMPOTENT_TOOLS (W113) are now built from the @_tool(read_only=..., idempotent=...) decorator kwargs instead of being hand-maintained sets that drift from the decorator truth. Only _CORE_TOOLS remains as a hard collapse candidate; _REGISTERED_TOOLS and _TOOL_METADATA are non-collapsible by design (the former is fastmcp's live registry, the latter is the import-time catalog the new derivations read from).
• CLAUDE.md Pattern 4 reworded to reflect the Fix-G resolution (W139). All 5 conventions sites (describe, understand, minimap, preflight, conventions) delegate to conventions_helper.compute_conventions(). --persist lives ONLY on the standalone conventions command, so the registry-write path stays single-owner.
• CLAUDE.md findings-registry section refreshed (post-W138). 14-detector list with the canonical confidence-tier vocabulary (static_analysis / structural / taint / heuristic), the expanded subject-kind vocabulary, and the version-stamp placement rule (stamps travel with the row, not the envelope).
• README.md headline reordered (W138) so the test_doc_consistency.py regex captures the canonical 149 MCP tools count instead of a transposed substring. Pure ordering change; numbers match roam surface --json exactly.
• llms-install.md swept for v13.0+ substrate (W124). 86 → 142 lines; added the 4 modes, the agent loop, the findings-registry section, and the 4 world-model classifiers. The file is now structurally equivalent to the long-form architecture doc but cropped to LLM-onboarding length.
• architecture.html findings-registry section added (W130). 142 lines documenting the registry schema, the 4 confidence tiers, the agent-loop step 7a (roam findings list between critique and pr-bundle emit), and the detector inventory with confidence-tier mappings.
• README + landing-page reframe — Option B "evidence compiler" (W171 / W178 proposals; W200 committed). Headline copy on README.md and templates/distribution/landing-page/ now leads with the evidence-compiler thesis (identity / authority / evidence) and positions the findings registry + Agent OS substrate as the inputs to a portable assurance packet, not standalone analyses.
• CLAUDE.md evidence-compiler thesis section + W148-doc adversarial fix (W170). Section 9 of the handover memo grew into a first-class CLAUDE.md block; the W148 doc-only fix for the adversarial recipe (registry-key lookup, not string-templated CLI invocation) is now reflected in the Pattern 5 commentary.
• CLAUDE.md cross-walk + identity/authority/evidence framing (W187). Added a phrase-level glossary so reviewers can map a finding/run/bundle back to the assurance packet without re-reading the architecture memo.
• policy_decisions promoted to typed PolicyDecision dataclass (W279 + W279b). New src/roam/evidence/policy.py mirrors the W211 ApprovalRecord pattern: frozen dataclass with explicit rule_id / decision / verdict / evaluated_at / extra, replacing the prior tuple-of-mapping shape. Two user-mandated integrity constraints landed under W279b: (1) PolicyDecision is a collections.abc.Mapping subclass so W226's apply_profile() _redact_mapping_tuple() keeps working unmodified; and (2) the legacy-preserve ValueError catch at change_evidence.py:320-358 was narrowed to only preserve rows when rule_id OR decision is missing, so drift-detection now correctly fires on {"rule_id":"r", "decision":"approved"} (a row that LOOKS valid but cannot construct a typed PolicyDecision). 31/31 golden content_hashes preserved.
• Trust-tier surface on roam evidence doctor (W281). New _classify_trust_tiers(packet) helper and extended _validate_closed_enums() check actor_refs[i].trust_tier membership against ACTOR_TRUST_TIERS. Doctor now always emits a Pattern-2 5-key trust_tiers dict (one entry per trust tier, zero where absent) plus a trust_warnings[] array. Verdict ladder extended: FAIL on enum violations or hash mismatch; WARN on PARTIAL / INSUFFICIENT banner OR on a STRONG banner with any self_reported_agent / unknown actor; PASS requires STRONG + zero trust warnings. Closes the silent pseudo-actor gap that the W278 classifier had only addressed at construction time.
• Pseudo-actor corroboration in classify_actor_trust_tier() (W285). The classifier gained two new kwargs: corroborated_tool_ids and corroborated_actor_ids, both frozenset[str] with exact-equality membership (no name allowlist, no substring match). A new collector helper _collect_corroborated_ids() at collector.py:1150-1357 reads two real evidence sources: HMAC-verified run-ledger events (mirroring cmd_runs._verify_one_run: ensure_ledger_key → read_run_events → verify_chain; only result["state"]=="ok" contributes, whole-run granularity) and parseable MCP receipts. _RUN_LEDGER_TOOL_FIELDS constant captures the non-uniform event-field naming for forward compatibility. As a side-effect this also closes the W197 bypass: MCP-receipt-mirrored ActorRefs that previously skipped the W278 classifier now flow through it. Real-world smoke on the roam-code workspace itself: doctor verdict WARN → PASS, with 3 previously-unknown-tier pseudo-actors (<unknown> / roam_init / roam_reindex) promoted to local_env via real HMAC-verified run-ledger corroboration. Negative proof: a tempfile.mkdtemp() workspace with no .roam/ keeps roam_init at unknown — no name-based shortcut, only real evidence promotes.
• Canonical demo evidence fixture repaired to true PASS (W286). The fixture's prior self_reported_agent actor (claude-code 1.2.3) was replaced with a local_env actor (example-trusted-agent) corroborated by an HMAC-signed run-ledger event, so the bracket test_doctor_passes_on_canonical_packet (canonical → PASS) vs the insufficient-banner fixture (→ WARN) now holds with real evidence on both sides. Cross-references updated: top-level agent_id, human_actor, approvals[0].approver, authority_refs[1].granted_by. Content hash recomputed via .audit-tmp/w286-rehash.py mirroring ChangeEvidence.compute_content_hash discipline.

> Evidence pipeline hardening batch (W279-W287 + W247a, > 2026-05-14). Typed PolicyDecision drift detection + packet > size budget + trust-tier surface + corroboration-based > promotion + provenance vocab + generated limitations + > producer-site version stamping + GitHub review parser. > Real-world roam evidence doctor on the roam-code workspace > itself: VERDICT PASS (was WARN via 3 unknown-tier > pseudo-actors; W285 promoted them to local_env via real > HMAC-verified run-ledger corroboration, with W286 confirming > the bracket against an insufficient-banner fixture). Real > packets sit at ~96 KB against a 256 KiB budget. Provenance > vocabulary landed clean; producer-side wiring is deferred to > W290+ so the call-site churn is decoupled from the vocab > freeze.

Research / planning

• MCP Pattern-1 family audit (W315). Sonnet+web research pass that named Variants A / B / C, surfaced the open Variant B gaps on doctor / stale-refs / triage / pytest-fixtures, and cited FastMCP / Anthropic / MCP-spec sources for the structured isError: true discipline. Memo lives at dev/MCP-PATTERN-1-FAMILY-AUDIT-2026-05-15.md. Drove the W325 <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> chokepoint fix and the W328 CLAUDE.md codification.
• Pattern 3 + Pattern 6 audit (W329). Sonnet+web pass that enumerated 9+ MCP parameter-name mismatches and 8 commands returning 50K-1.6M tokens with no auto-handle. Memo lives at dev/PATTERN-3-AND-6-AUDIT-2026-05-15.md. Drove the W330 <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> CLAUDE.md expansion, the W331 / W331b definition-field rollout, and the W332 input_path normalization.
• MCP state-mutating patterns (W340). Sonnet+web pass that catalogued every state-mutating MCP wrapper and surfaced 4 sub-waves (W363-W366) for hardening the mutation surface. Memo lives at dev/MCP-STATE-MUTATING-PATTERNS-2026-05-15.md. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• Standards currency audit (W341). Sonnet+web pass against SLSA v1.2, OSCAL v1.2, and the EU AI Act drafts. Surfaced 3 sub-waves (W358-W360) to refresh the in-tree control-map language. Memo lives at dev/STANDARDS-CURRENCY-AUDIT-2026-05-15.md. Headline finding: <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> SLSA v1.2 Source Track + OSCAL v1.2 Control Mapping landed after the W184 control-map shipped, so the in-tree wording ("maps to" / "supports evidence for") still holds but the cited revision numbers need a refresh pass.
• Detector competitive audit (W368). Sonnet+web pass against the open-source detector landscape (Snyk / Sonatype / Semgrep / SonarSource). AHEAD on 5 categories (vibe-check/AI-rot, agentic audit-trail, taint+OpenVEX, N+1 implicit-property, graph-coupled bus-factor); PARITY on 5; BEHIND on 6. Top 3 gaps: reachable-vuln open-source parity, smells rule depth (empty-catch + primitive-obsession), taint OWASP Top-10 sink coverage. 5 sub-waves queued (W370-W374). Memo lives at dev/DETECTOR-COMPETITIVE-AUDIT-2026-05-15.md. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• Ecosystem positioning audit (W385). Sonnet+web pass against 7 adjacent tools (review tools, IDE plugins, governance platforms, supply-chain analysers, agent observability). Grades each on the agentic-assurance axis: 5 COMPLEMENTARY (different surface; integrate-not-compete) / 2 COMPETITIVE (overlapping evidence claim) / 0 SUBSTITUTE (no full replacement). Confirms the "local evidence compiler" thesis — no surveyed tool emits a portable evidence packet, so Roam's compiler position is defensible. Memo lives at dev/ECOSYSTEM-POSITIONING-2026-05-15.md. Feeds the W390 / W393 <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> pitch-refresh wording and the W397 auto-count Codex-headline template update.
• Performance benchmarking research (W395). Sonnet+web pass against 7 indexing tools (ctags, jedi, aider repo-map, serena, Sourcegraph SCIP, CodeQL, tree-sitter raw). Positions roam at MEDIUM: 10-40x slower than ctags per-file (different category — ctags emits tags only, no graph / git / metrics) but 5-20x faster than CodeQL for comparable analysis depth on a typical 50K-LOC repo. Defensible claim: "Roam indexes a fresh 50K-LOC codebase in 15-30s and re-indexes a single changed file in under 5s." Recommends a 5-wave optimization sequence — W400 phase-timing in roam doctor (prereq), W404 (≡ W396 in the memo) ROAM_PARALLEL_INDEX default-on, W397-equivalent shallow-git default on first index, W399 Louvain cache expansion, W398-style parallel parse via ProcessPoolExecutor. Memo lives at dev/PERFORMANCE-BENCHMARKING-2026-05-15.md. The five sub-waves <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> are queued as W404 + W405-W408 in the backlog (renumbered to avoid collision with the in-tree W396-W400 series).
• llm-smells pattern catalog (W402-research). Sonnet+web pass that designs roam llm-smells, a NEW command detecting anti-patterns in HUMAN code that calls LLM APIs (openai, anthropic, google-generativeai, langchain, litellm, etc.). Distinct from vibe-check (which detects AI-generated code shape). Catalog has 14 patterns: 8 CHEAP (regex-based; ship in v1), 3 MODERATE (use existing symbols.default_value column), 3 EXPENSIVE (deferred — need new dataflow edge types). v1 surface = 11 patterns across token budgets, model pinning, system messages, error handling, retry discipline, and context-window assumptions. Primary academic source: Mahmoudi et al. arXiv:2512.18020 (Dec 2025), 200-system empirical study with 86.06% average detection precision. Memo lives at dev/LLM-SMELLS-PATTERN-CATALOG-2026-05-15.md. Estimated v1 <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> implementation: 6-8 hours (core detector + 11 patterns + MCP wrapper + tests). When shipped this is the first production-grade multi-provider linter for LLM API anti-patterns.
• OWASP 2026 taint rule pack research (W372-research). Sonnet+web pass that surveyed the OWASP 2026 top-10 against the current in-tree taint rule set. Memo at dev/OWASP-TAINT-RULE-PACK-RESEARCH-2026-05-15.md. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> Identified 3 first-ship rules ranked by precision-vs-effort: W373 (python-ssti — Jinja2 server-side template injection), W374 (java-sqli — JDBC string-concat SQL injection), W375 (java-deserialization — ObjectInputStream deserialization sinks). All three rules use existing AST edge types; no dataflow-engine extension required. v1 estimated implementation: 3-5 hours per rule.
• Phase 4-7 perf research (W395-followup). Sonnet+web pass that ranked the remaining phases (4-7) of the W395 perf roadmap against real-world signal. Memo at dev/PERF-PHASES-4-7-RESEARCH-2026-05-15.md. Top 3 actionable <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> outcomes: W423 (PageRank warm-start; 2-5s savings on warm re-index), W424 (SQLite synchronous=NORMAL pragma; 1-3s savings on full-index commits), and W407 reclassified to VALIDATE — the Louvain cache is already implemented; the memo's recommendation was to verify (not build). NOTE: this ranking was superseded by the W408 real-data finding (effects_taint consuming 48% of indexer wallclock); the new ranking pushes W433 (effects_taint optimization) ahead of W423.
• Standards crosswalk additions research (W360-research). Sonnet+web pass that surveyed standards-update activity since the W341 audit. Memo at dev/CROSSWALK-ADDITIONS-2026-05-15.md. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> Proposes 5 new YAML entries: 4 NIST AI 600-1 controls + 1 NIST SP 800-218A control. CAISI (US AI Safety Institute) is held until H2 2026 — its standards remain in concept-paper form and adding citations would invite later wording-drift. Implementation queued as W428.

Added — taint rule pack v1 + shallow git default + auto-generated MCP table + oracle wrapper dedup + qualified-name flag (W466 batch)

• W405 — shallow git history default on first index. git_stats.py now defaults to a 365-day window via _DEFAULT_SINCE; cmd_init.py surfaces --full-history for the opt-out and ROAM_GIT_SINCE for env override. The _first_index() gate preserves existing deep indexes so existing users see no behaviour change on re-index. 30 + 31 + 115 focused tests pass. Drive-bys W437/W438/W439 queued. Lowest-risk of the W395 perf sub-waves.
• W373 — python-ssti taint rule (T-X01, CWE-94). Server-Side Template Injection rule against the existing taint engine — engine already supports qualified-name matching, so no engine extension required. 7 new + 45+39 existing tests pass. Drive-bys W452/W453/W454 queued (W454 also landed this batch).
• W374 — java-sqli taint rule (CWE-89). Java SQL Injection rule with the same recall-limited precision profile as java-fileupload because the engine lacks Java qualified-name resolution today. 7 new + 44+31 existing tests pass. Drive-bys W455/W456/W457 queued — W455 captures the "engine needs Java qualified-name resolution" follow-up surfaced during this wave.
• W454 — per-rule qualified_only flag for taint engine. Lets individual rules opt in to qualified-name-only matching; java-sqli opts in so the recall-limited precision profile is per-rule, not engine-wide. 29 + 60 focused tests pass. Drive-bys W461/W462/W463 queued.
• W443 — README coverage for 4 untracked CLI commands. Added evidence-diff, evidence-doctor, llm-smells, and findings to the README command index. test_readme_covers_all_canonical_cli_commands now passes. Drive-bys W449/W450 queued.
• W449 — auto-generated README MCP tool table. New surface_counts.mcp_tool_descriptions() helper drives a generated table; 74 missing tools added and the core preset count corrected (25 → 57) so it matches roam surface --json. Closes the long-standing 73-tool drift surface that the W411 backstop wave had only patched. 4/4 + 16/16 + 8/8 + 31/31 test suites pass. Drive-bys W458/W459/W460 queued.
• W432 — oracle wrapper dedup. All 5 oracle wrappers (symbol_exists, route_exists, is_test_only, is_reachable_from_entry, is_clone_of) had been duplicated by W306; this wave removed the duplicates. Decorations went 228 → 223 unique, which matches the CLAUDE.md headline (mcp tools registered: 223). Added a new AST duplicate-name CI lint via surface_counts.mcp_tool_decorations() so this class of drift fails at lint time. Drive-bys W443/W444/W445 queued (W443 landed this batch).
• W429 — small-cleanup bundle (W422 + W425 + W426). Three drive-bys from the W383 + W421 batch landed together: W422 marks the standalone permit wrapper as deprecated, W425 adds lease warnings_out, W426 surfaces a constitution-unparseable warning. 204/204 tests pass; 31/31 hash stability byte-identical. Drive-bys W446/W447/W448 queued.

Research / planning — W466 batch

• W433-research — effects_taint optimization scoping. Memo at dev/EFFECTS-TAINT-PERF-RESEARCH-2026-05-15.md. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> Three candidates ranked: (C) double-parse I/O elimination — 15-30s, zero risk; (B) function-summary memoization 35→5s; (A) file-signature cache warm-reindex 0s on cold path. Surprise discovery: roam has TWO independent taint engines — analysis/taint.py for Phase 5 (indexer-side) vs security/taint_engine.py for the roam taint command. Consolidation is a deeper structural play beyond the immediate W433 perf wave.
• W358-research — SLSA v1.2 Source Track positioning. Memo at dev/SLSA-V12-POSITIONING-2026-05-15.md. roam de-facto covers SRC-L2 today. SURPRISE: SRC-L3 lift is ONE wave — cosign_sign_statement() at attest/cga.py:495-594 is already implemented. New wave W451 queued to formalize the SRC-L3 mapping. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. -->
• W359-research — OSCAL v1.2 Control Mapping decision. Memo at dev/OSCAL-V12-CONTROL-MAPPING-2026-05-15.md. <!-- PHANTOM 2026-05-18: declared SHIPPED but memo absent from disk. Regenerate from BACKLOG/test-fixture breadcrumbs before next release. --> SURPRISE: OSCAL v1.2 added a 7th model (Control Mapping) which is the zero-prerequisite first emission for per-run evidence. AR for per-run evidence. New waves W464/W465 queued.

Added — permit unification + Pattern-3b extension + llm-smells v1.1 + phase timing

• W377-batch — six permit-persist red-team gap closures (W377-W382). Sealed the six drive-by gaps W349 surfaced against the W198 permit writer. Touched src/roam/commands/cmd_pr_bundle.py _load_permits_from_disk and src/roam/evidence/collector.py _build_authority_refs. 31/31 golden hashes byte-identical (the W292/W294 authority-axis stability proof held); 163 focused tests pass. No ChangeEvidence.content_hash movement.
• W383 — unified permit reader behind canonical store helper. roam.permits.store.load_permits_from_disk is now the single canonical reader; both cmd_pr_bundle and cmd_pr_replay delegate. Closes the historical split-brain where the two surfaces applied independent (and slightly different) JSON-parse + expiry filtering. 163/163 focused tests pass; 31/31 golden hashes remain byte-identical. Two drive-bys captured as W421/W422 for follow-up.
• W347 — Pattern-3b parameter-alias normalization extended. Added file_path → path to _PARAM_ALIASES in src/roam/mcp_server.py. The prefix-pattern cluster (queries/prefix/patterns) was deliberately bailed on this wave — the boundary normalization shape doesn't fit a clean one-name canonical without consumer-side churn. 2733 + 140 + 31 focused tests pass. Three drive-bys queued as W430/W431/W432.
• llm-smells v1.1.0 — five new CHEAP detectors (W415b). Five additional regex-based detectors landed against the v1 surface: missing_timeout, missing_max_retries, no_system_message, no_retry_backoff, call_in_loop. Total v1.1 detector count = 10. 36/36 focused tests pass. Package version bumped 1.0.0 → 1.1.0. Three drive-bys queued as W415c/W415d/W427 for next-wave follow-up (additional patterns + MCP wrapper polish).
• W408 — per-phase timing instrumentation in roam doctor. Indexer phases (discover / parse / extract / resolve / effects / taint / graph / metrics) now emit per-phase wall-clock timings surfaced in roam doctor. Real-data finding (CRITICAL): on roam-code itself, effects_taint is 48% of indexer wallclock (67.6s of 139.6s). This invalidates the PageRank-first ranking in the W395-followup memo: the new ordering pushes W433 (effects_taint optimization) ahead of W423 (PageRank warm-start) and W424 (SQLite pragma). 134/134 focused tests pass. Three drive-bys queued as W433/W434/W435.

Bailed

• W421 — constitution + lease gatherer delegation audit (BAILED). Investigation found that cmd_pr_replay constitution + lease gatherers already delegate to the canonical readers; no refactoring was warranted. 119/119 baseline tests pass. Two drive-bys captured as W425/W426 (informational follow-ups, not fixes).

Fixed

• roam taint engine BFS no longer silently returns 0 findings (latent PRE-v13 regression — see Added section above).
• roam doctor 17 → 20 checks documentation drift corrected.
• db/connection.py:255 missing Callable import (F821) — type annotation would have raised NameError under get_type_hints evaluation.
• world_model/tx_boundaries.py:474 confidence = "medium" unused local — classifier had been under-reporting medium-confidence transaction boundaries.
• test_neighbor_hub_files_are_dropped + test_low_degree_neighbor_still_expands stale .gitignore-interaction fixture (Case A — test isolation).
• test_batch_mcp mock signature stale after the include_paths kwarg addition — fix surfaced a real production coarse-try/except swallowing parameter mismatches.
• 2 SyntaxWarning raw-string fixes: \P in laravel_post.py:16 and test_atomic_io_consolidation.py.
• batch_search / batch_get per-query exception handling (W103). The coarse outer try/except in both MCP batch wrappers aborted the entire batch on a single per-query exception — one bad symbol-id in a 50-symbol batch would zero-out the other 49 results. Restructured so the outer try/except protects only open_db(); each per-query body has its own try/except inside the loop, surfacing per-row error fields without killing the batch. Surfaced during W101 test investigation as a real production bug, not a test artefact.
• stale-refs --attest <PATH> mkdir crash (W126). The unprotected Path(attest).parent.mkdir(parents=True) propagated FileExistsError / NotADirectoryError / PermissionError, aborting the entire stale-refs scan before any finding was emitted. Now wrapped in try/except OSError; surfaces summary.attest_error + summary.attest_status while the scan completes. Exit-code policy: 0 without --gate, 6 (PARTIAL) with --gate when there are no other findings, 5 (GATE_FAILURE) when other findings exist. Was a HIGH-severity finding in W112.
• cmd_surface --json returns env-independent mcp_tool_count (W138). Previously returned 0 when fastmcp was not installed; now reads _TOOL_METADATA so the inventory matches the documented count on every install variant.
• Lease midnight-UTC clock race (W135). tests/test_lease_system.py previously used a ±1d tolerance to absorb test runs that crossed midnight UTC; the test now freezes the clock via monkeypatch on roam.leases.store._utc_now* so the assertion is exact. W112 medium-severity finding.
• vibe-check tier-mapping discipline (W125). empty_handlers and abandoned_stubs were originally tagged static_analysis but the detector is regex-based, not AST-based — both downgraded to heuristic. hallucinated_imports upgraded to structural (it uses graph reachability, not pattern matching). Codified the rule: static_analysis is reserved for deterministic AST/CFG plus taint or dataflow plus cross-reference checks, and is never assigned to regex name matching.
• _check_index_step_failures doctor advisory verified working through W127 — no behaviour change this session, but the W82-era advisory was re-confirmed against the new init-time cloud-sync warning to ensure they do not double-fire on the same root cause.
• PR Replay Markdown output — 3 hostile-input bugs (W217). New hostile-input snapshot tests caught real broken-output paths: pipe characters inside a code field broke the surrounding Markdown table; embedded newlines in summary.verdict strings unwrapped onto the following row; backtick-bearing identifiers escaped the inline code span and ran into the next sentence. All three sealed with per-cell escaping at the renderer; tests now pin the byte-output shape against fixture inputs.
• W184 control-map wording drift — 3 entries (W203). The CI wording-guard lint surfaced three control-map entries that overclaimed certification ("certifies", "makes compliant" — instead of "maps to" / "supports evidence for"). All three rephrased in place; lint now blocks future drift.

Security

• roam_annotate_symbol silent-success fix (W324). The MCP wrapper had been returning verdict: "completed" even when the underlying resolver degraded to a no-match — a Pattern-2 violation (silent fallback). The wrapper now surfaces state: "no_match" / verdict: "no annotations applied" with a next_command pointing at roam search when resolution degrades, so the LLM sees the failure mode instead of acting on a false-PASS. This was the surfacing event that produced Pattern-1 Variant D (W334).
• cmd_impact weighted_impact rounding + silent-fallback fix (W336). Two Pattern-2 violations sealed in one wave: (a) the weighted-impact field was being rounded to 2 decimals before the threshold comparison, so the verdict could silently flip across a rounding boundary; (b) when the upstream graph traversal returned zero edges, the command emitted verdict: "low impact" instead of state: "no_callers_found". Both paths now route through the always-emit-state Pattern-2 contract, and the threshold comparison uses the un-rounded weighted-impact value.
• Redaction-leak snapshot tests (W232). Five passing redaction snapshots pin the existing behaviour; four known leak paths surfaced and pinned with pytest.mark.xfail(strict=True) so they auto-clean as fixes ship — the test goes RED if a leak is accidentally sealed without removing the xfail. The four leak paths are: (1) absolute filesystem paths in EvidenceArtifact.path when the artifact is outside .roam/; (2) MCP receipt args carrying secret-looking strings that the redactor's allowlist misses; (3) git-author email leaking into actor_refs when trust_tier == git_author; (4) constitution-hash leaking the unredacted constitution YAML when constitution_hash is computed on a permit-bearing constitution.
• Schema migration golden tests — hash stability proven byte-identical across W174 → W182 → W210 (W218). 5 fixtures (minimal, ideal-case, redaction-heavy, multi-actor, multi-authority); each round-trips the same content hash from the W174 baseline through W182's decision-receipt addition and W210's 9-field extension, so the backward-compat policy (_W210_OMIT_WHEN_DEFAULT_FIELDS) is now a proven invariant, not an aspirational one.
• Producer/collector contract tests — 24 tests, 5 producer gaps pinned (W219). End-to-end pairs (producer emits → collector consumes) for each of the 6/8 wired evidence slots. The 5 still-open producer gaps are now xfail-strict so they fail loudly the moment the producer side ships.
• Executable 8-question audit (W220). Threshold-gated test that exercises the canonical 8/8 ideal-case fixture end-to-end and asserts every assurance question (Who? What authority? What was read? What changed? What broke? What was approved? What was redacted? What evidence is stale?) is answerable from the packet alone.
• Layer-2 collector secret scrub (W249). Defence-in-depth pass in src/roam/evidence/collector.py: new _scrub_actor_block helper reuses W241's _redact_secrets_in_string verbatim and fires on 8 envelope fields — top-level verdict + summary.verdict + the actor block (actor.agent_id / agent / human_actor / human / user / display_name / mcp_client_id / tool_id / ci_runner_id) + legacy top-level agent_id / human_actor. Boundary re-scrub at _build_actor_refs provides a second line. "secret" stamps in bundle_redactions are deduped against producer-stamped W240 entries. Both W232 xfail-strict tests flipped to PASS — markers removed from tests/test_evidence_redaction_snapshots.py. Justified by three real failure modes: pre-W240 envelopes on disk, third-party producers that bypass pr-bundle, test fixtures feeding dicts directly into the collector.
• W219 gap-pin verification (W255). 4/5 originally-pinned producer-side contract gaps cleanly flipped GREEN: pr-bundle.context_files (W224a), pr-bundle.approvals / accepted_risks (W224a + W240 CLI), pr-risk.findings[] (W242). The 5th (pr-bundle.mode) was already producer-side sealed by W224c but its contract test docstring was stale — sealed by W257. The tool_id reservation (no test) stays as a None placeholder for future W196 emitter work.
• pr-bundle.mode contract test refresh (W257). Updated tests/test_producer_collector_contracts.py::test_pr_bundle_mode_contract (lines 334-402): docstring rewritten to reflect the post-W224c reality (producer ALWAYS emits mode + summary.active_mode, defaulting to "unmoded"); producer-side assertions added ("mode" in envelope; envelope["mode"] in VALID_MODES | {"unmoded"}; summary.active_mode mirrors mode). Collector-side safe_edit injection check preserved. 24/24 contract tests pass.
• Critique-contract drift-guard pin (W256). 14-line constant-level test in tests/test_evidence_collector.py:1504+ pins 'check' in _FINDING_SAFE_KEYS — the field the critique envelope carries through to ChangeEvidence.findings[]. The pin is intentionally scoped to the one constant that produced observed drift; speculative pins on rule_id / redactions were rejected to keep the guard focused. Prevents a silent regression where a _FINDING_SAFE_KEYS rewrite drops the critique check field and the collector starts emitting partially-populated critique findings without test failure.
• Q8 producer_not_available redaction route (W261). Extended REDACTION_REASONS 8 → 9 (new member producer_not_available, declared in src/roam/evidence/_vocabulary.py); no schema expansion — the existing redactions tuple was reused. PR Replay emits a Q8 limitation entry at src/roam/commands/cmd_pr_replay.py:1212-1242 whenever the packet contains no approvals (forward-compatible: the marker silently disappears once a real approvals producer ships, no consumer change required). Renderer dedicated Q8 bullet at cmd_pr_replay.py:1797-1827. Audit harness gained an asymmetric EXPECTED_PARTIAL_COUNT_TODAY = 1 slot. Real-world smoke on roam-code: complete=7 partial=1 missing=0; honest-banner tier stays STRONG. Closes Q8 as the last silent gap on the pipeline-coverage arc.

Performance

• roam n1 bulk-fetch deeper helpers — 2 new bulks for $with resolution + resource-config; ~200 SQL roundtrips + ~500 disk reads saved per invocation on 100-model Laravel applications.
• roam n1 candidate-filter N+1 closed — gap-models filter collapsed from 100 queries to 1 (batched_in).
• roam index git-history skip when HEAD is unchanged — 94% speedup on warm re-runs (sentinels added in W87; original ship was pre-13.0).
• FTS5 docstring column with BM25 weight 4 (sentinels added in W94; original ship was pre-13.0). USER_VERSION discipline widened to hash _FTS5_SCHEMA_COLUMNS (W97).
• CLAUDE.md auto-count generator now produces accurate counts in CLAUDE.md, README.md, and llms-install.md (W138). dev/build_readme_counts.py:_claude_blocks was emitting the hand-stale 200+ commands / 130+ MCP tools placeholder; it now emits 234 commands · 149 MCP tools (57 in the default \core\ preset) · … driven off roam surface --json, so the three doc surfaces converge on the same number on every regeneration.

Deprecated

• (Already shipped in 13.0) 7 redundant aliases live in _DEPRECATED_COMMANDS in src/roam/cli.py. W66 backfilled the missing 13.0 CHANGELOG entry.

[13.0] — 2026-05-13

Deprecated

• Seven legacy aliases now emit a deprecation note (BACKLOG Rank 18 / W3.3). digest/math/refs/snapshot/trend/onboard/churn graduated out of the _INTENTIONALLY_UNCATEGORISED allowlist in tests/test_surface_consistency.py into _DEPRECATED_COMMANDS in src/roam/cli.py. Each alias still resolves to its canonical command (trends/algo/uses/trends/trends/understand/ weather) — this is informational, not breaking — but every invocation now prints DEPRECATION: '<alias>' is an alias for '<canonical>' … on stderr and, in --json mode, carries the same string under summary.deprecation_warning so JSON-only consumers (CI, MCP clients) can detect the rename mechanically. Removal target is unset; users have at least one release cycle to migrate. Contract pinned by tests/test_alias_deprecation.py (5 tests).

Substrate evolution

• R20 ledger HMAC chain with CGA signing
• R21 multi-agent lease system
• R25 plugin substrate validated end-to-end via Rails Path A clean cut
• atomic_io consolidation across writers
• Constitution/policy unified single source of truth
• World model: side-effects, idempotency, causal graph, tx-boundaries classifiers

Mode enforcement (staged rollout, opt-in)

• _MODE_ALWAYS_ALLOWED bootstrap allowlist (index/init/reindex)
• Test fixture sweep across 10 privileged-command test files
• Classification coverage for 8 additional commands
• Empirically validated: 251/252 pass with ROAM_MODE_ENFORCEMENT=1

Real-world feedback fixes

• stale-refs heading-slugger now matches GitHub's algorithm (underscore + whitespace)
• stale-refs --fix corruption guards: URL-half resolution + bare-backtick suppression
• algo nested-lookup dataflow predicate + PHP ===/!== detection
• auth-gaps helper indirection (2-level same-class + ancestor descent)
• 7 of 8 Laravel dynamic-dispatch idioms detected (Route/Eloquent-scope/Policy/ Observer/Job/Queue/Artisan) via new roam/index/laravel_post.py
• over-fetch 3-state classification (BARE / GUARDED_RELATION / UNGUARDED_RELATION)
• pr-bundle --strict-resolved flag + --ci global mode integration
• ws resolve exposes unmatched URLs with reason classification

New commands

• roam brief, roam next, roam mode, roam constitution, roam laws
• roam memory, roam lease, roam runs, roam replay, roam agent-score
• roam agents-md, roam architecture-drift, roam graph-diff
• roam side-effects, roam idempotency, roam causal-graph, roam tx-boundaries
• roam batch-search, roam complete, roam mcp

Drift-guard infrastructure

• LAW 4 anchor lint + AST-pinned CLAUDE.md count assertions
• Capability registry auto-derive (eliminated 531 LOC of bookkeeping)
• README.md marker-based count auto-generation + doc-hygiene CI workflow
• --sarif consumer drift-guard + --budget coverage survey
• Canonical-constant citation lint (load-bearing, allowlist empty)
• Extension-constant consolidation (DOC_EXTENSIONS canonical; _BINARY_EXTENSIONS aliased to SKIP_EXTENSIONS)
• Anti-leak CI gate extended with domain-term patterns

Renamed

• summary_envelope → strip_list_payloads (name reflects contract)

Schema

• USER_VERSION 12 → 13 (migration #51: loop_eq_with_dependent_write column for algo nested-lookup dataflow predicate)

Round 10/11 — 9-agent parallel hardening pass

Nine Opus agents in two parallel waves closed the gaps surfaced by the R9 cross-rechecks. All landed in the working tree; none lost when the host PC died mid-flight.

R10 wave — 5 agents

• R10.1 — roam mutate --apply rollback bug fixed + apply-mode tests for all 4 transforms. _apply_move had no rollback: a mid-flight OSError left the destination file with the moved symbol AND the source file still containing it (duplicate definitions in the repo, raw OSError propagated up). Snapshot-and-restore now buffers pre-apply state of every touched file; on failure, newly-created targets are removed and pre-existing files are rewritten from snapshot. Wraps the exception into {isError: True, error_code: "APPLY_FAILED", files_modified: []}. New tests/test_mutate_apply.py adds 10 tests across rename / add-call / extract / move-rollback paths — all assert on file bytes after --apply, not just dry-run JSON shape (the prior gap).
• R10.2 — Handle-off cache GC + SBOM as a Release asset. .roam/responses/ used to grow unbounded (disk DoS + forensic leak). New _gc_handle_dir() runs amortised (once per 25 writes or when dir > 50 files): TTL eviction (ROAM_MCP_HANDLE_TTL_HOURS, default 168 = 7d) plus oldest-mtime LRU under a max-bytes cap (ROAM_MCP_HANDLE_MAX_BYTES, default 100MB). Race-tolerant per-entry try/except, 0o700 dir mode, defence-in-depth parent.resolve() != handle_dir.resolve() guard. SBOM is now uploaded to the GitHub Release on tag-triggered runs via gh release upload --clobber, cosign-signed (keyless OIDC, same Sigstore chain as cga emit) — procurement no longer has to dig through expired workflow artifacts.
• R10.3 — Louvain cluster cache + roam_context bulk-fetch. Cluster phase on a no-op reindex: ~4.3s → 63ms (signature {n, m, top-64-degree-ids} persisted in index_manifest.notes; _run_clustering skips when signature matches and clusters table is non-empty; force=True bypasses). roam context <symbol> query count: 1,986 → 39 (well under the 50 ceiling). get_blast_radius, get_affected_tests_bfs, and get_coupling in context_helpers.py now bulk-load reverse adjacency once; the BFS is in-memory after that. Total roam index wall on this repo: 100s → 93s on no-op.
• R10.4 — roam_validate_plan warning-branch coverage. 12 new tests pin every WARNING code path: NAME_COLLISION, MEDIUM_BLAST_RADIUS, HIGH_BLAST_RADIUS, FITNESS_VIOLATIONS, INVALID_TARGET_FILE, plus aggregation (3 ops × 3 warnings → warnings_count == 3) and verdict precedence (blocker dominates warning). The agent also surfaced that FITNESS_VIOLATIONS was effectively dead code against real preflight envelopes — fixed by R11.C.
• R10.5 — Language-extractor smoke matrix. Eight extractors with no symbol-extraction tests now have one each via a parametrised matrix in tests/test_extractor_smoke.py: Apex, Aura, Visualforce, SFXML, HCL, Swift, FoxPro, generic. Each verifies module import + ≥1 symbol + canonical entity present + extract_references returns a list. tree-sitter-language-pack updates that silently break Salesforce/Apex (a real enterprise wedge) now show up in CI.

R11 wave — 4 agents

• R11.A — Three more N+1 fixes (cmd_attest, cmd_module, cmd_dead). _collect_blast_radius in cmd_attest.py:158 was issuing one SELECT FROM file_edges per source file (50-file PR = 50 queries; 500-file enterprise PR = 500). cmd_module.py:_module_deps and _collect_sym_ids ran for fid in file_ids: conn.execute(FILE_IMPORTS, (fid,)) — 100 queries per roam module <dir> on a 50-file directory. cmd_dead.py:_predict_extinction issued a SELECT per BFS pop (200-node fragment = 200 queries). All three rewritten with batched_in + a single reverse-adjacency dict; CountingConn regression tests pin constant-query behaviour up to n=50. 5-100× speedup depending on input size.
• R11.B — format_table single-pass. Stringify each cell once during the width pass and reuse during emit (instead of len(str(cell)) twice). 200 rows × 6 cols: 0.99 ms → 0.86 ms (~1.15×). Modest because table rendering is intrinsically O(rows × cols) — the win is from removing the redundant str()/len() pass. 6 byte-identity tests pin output equivalence with the pre-refactor reference.
• R11.C — Contract drifts (3 fixes). cmd_preflight.py now populates summary['fitness_violations'] as a list (was at r['fitness']['rule_details'] only — _vp_validate_one's warning was effectively dead). APPLY_FAILED (introduced by R10.1) added to _DOC_LINKS + _SEVERITY_MAP. tests/test_mcp_server.py _CORE_TOOLS count drift fixed: 51 → 57 (R8 added roam_validate_plan + 4 roam_for_* + roam_fetch_handle; expected set extended). Cascading update to test_multiple_warnings_aggregate_in_summary since FITNESS_VIOLATIONS now actually fires.
• R11.D — Help-template ratchet 32 → 52 commands. Twenty more commands pinned: plan, plan-refactor, suggest-refactoring, partition, layers, cut, orchestrate, agent-plan, agent-context, pr-diff, why, intent, capabilities, capsule, trends, audit, bus-factor, effects, grep, ci-setup. 156 generated tests (52 × 3) all green. Cross-references verified against _COMMANDS.

Round verification: 624 targeted tests across the 15 touched test files pass on first run after the crash recovery — zero seam issues across the 3 files that received edits from 2 agents (mcp_server.py, test_validate_plan.py, test_n1_fixes.py).

Round 9 — code work + 5-pass cross-recheck

• R9.B7 — FTS5 incremental sync. build_fts_index now diffs symbol_fts against symbols and only INSERTs new rowids / DELETEs stale ones instead of full DELETE+INSERT on every reindex. Synthetic 5,000-symbol benchmark: cold build 157ms, no-op incremental 15ms (10.5× speedup), 100-row diff 16ms (10.7×). Force-rebuild path retained for roam index --rebuild. New tests/test_fts5_incremental.py (5 tests) pins the convergence + cost properties.
• R9.A5 — Extracted baseline-diff branch from health(). cmd_health.py:health() was 920 lines; the if baseline_ref: branch (125 lines) is now a dedicated _emit_baseline_diff helper. Function down to 822 lines. Snapshot tests pass byte-identical.
• R9.A2 — Numbered migration ledger. _MIGRATIONS = [(seq, name, fn), ...] in db/connection.py is now the source of truth for schema migrations. MIGRATION_OPS_COUNT = len(_MIGRATIONS) (auto-derived); the previous hand-pinned constant was off by 5 (counted comment occurrences). New tests pin (a) the derivation, (b) seq monotonicity, (c) idempotency on a fresh DB.
• R9.G9+ — Help-template ratchet extended to 32 commands. Added tour, file, uses, trace, deps, health, diagnose, complexity, pr-risk, affected-tests, fan, hotspots, n1, clones, simulate, mutate, index, watch, surface, taint, vuln-reach, attest (delegated to a parallel agent for the mechanical sweep). Each gets \b + Examples: + See also block. 96 generated tests pass.

Recheck-driven fixes (R9 cross-pass)

• _strip_url_credentials URL-bug fix. Previous implementation used rpartition("@") over the whole post-:// string, which finds the LAST @ anywhere. URLs with @ in the path or query (e.g. ?reviewer=a@b.com, @scope/pkg paths) got rewritten to the wrong host, putting incorrect subject.name in every signed CGA. Fix: scope the credential strip to the authority slice only (per RFC 3986 §3). Regression test added.
• _DOC_LINKS + _SEVERITY_MAP plug 7 missing error codes. EMPTY_INPUT, INVALID_DIFF, RUN_FAILED, JSON_DECODE, ELICITATION_REQUIRED, FILE_NOT_FOUND, DIRTY_TREE were emitted in production but absent from both maps — agents got generic UNKNOWN doc_links and default error severity for diff/critique paths. Both maps now carry every code in use.
• Error-storm trim preserves retryable + doc_link. When the same error_code fires ≥3× the trimmed envelope used to drop both fields. Agents that branch on retryable (DB_LOCKED, INDEX_STALE) silently stopped retrying after 3 fires. Both fields are now always carried in the trimmed shape.
• PRAGMA busy_timeout=30000 pinned explicitly. Python's sqlite3.connect(timeout=30) sets the driver-level retry budget but not the engine's busy_timeout PRAGMA. Raw-sqlite3 consumers (test fixtures, MCP test paths) now see the same 30s budget as open_db.
• cmd_pr_risk early-return paths now wrap with json_envelope(). Two --json exits previously emitted bare dicts (no schema_version / no summary.verdict), breaking downstream consumers expecting the contract.
• Surface-count drift cleaned up across docs. 137 MCP tools → 145, 122 tools → 145, 5 prompts → 6, 25 default core tools → 58, 208 commands → 211 (or 204 + 7 aliases) — applied to README.md, CLAUDE.md, landing-page index.html, docs/index.html, docs/command-reference.html. CLAUDE.md Documentation Hub also re-pointed at the canonical templates/distribution/landing-page/docs/ path (the deleted docs/site/ was still referenced).
• Broken cookbook link removed from docs/command-reference.html — docs/cookbook/ was deleted in 12.50; replaced with a roam ask --list reference.

Correctness

• Pure renames now recover affected-neighbor edges. indexer.py:1409 had if not force and modified and changed_file_ids — pure renames produce modified=[] and silently skipped neighbor recovery, leaving roam impact <renamed_symbol> reporting fewer callers than reality. Drop the spurious and modified clause; add tests/test_index.py::test_pure_removal_invokes_affected_neighbor_recovery as a spy-based regression guard.
• roam cga verify fails closed when no .bundle is present. Previously: a downloaded statement without its sibling bundle reported verified while cosign: null — the cryptographic-trust half was silently skipped. Now: refuses unless --no-cosign is passed to acknowledge predicate-only verification.
• git_dirty_hash + git_commit_sha1 bound into predicate verification. Predicate now carries git_dirty_hash; verifier reads live values and refuses on mismatch (clean-vs-dirty, dirty-hash drift, commit-SHA mismatch). roam cga emit refuses on a dirty working tree by default — --allow-dirty records the dirty-hash and proceeds.
• SQLite USER_VERSION bumped 1 → 12 with a discipline pin. New MIGRATION_OPS_COUNT constant + tests/test_db_user_version.py fail CI when ensure_schema gains a migration without a corresponding bump. Closes the loop on the existing index_manifest.schema_version writer.

Performance

• N+1 fixes (4 sites). _find_colocated_tests (3 nested N+1 → 1 bulk fetch), _print_mega_detail (per-cluster → batched IN), _against_mode (per-fid co-change query → batched), _find_eager_loads (controller-file cache, 5-10× speedup on Laravel roam n1).
• N+1 fixes — roam n1 deep pass (4 more sites). _find_appends_properties, _find_accessor_methods, _trace_io_via_edges, _find_collection_contexts now drive off pre-loop batched_in fetches in analyze_n1 instead of running 1-3 queries per model. tests/test_n1_fixes.py::test_analyze_n1_appends_collection_edges_constant_query_count pins constant-query behaviour from 1 to 50 models. Old per-loop overhead on a 100-model app: ~5 × 100 = 500 queries; now: 4-6 batched queries total.
• Inline phase progress on roam index. Adds [1/7] … [7/7] markers before each pipeline phase (parse → resolve → graph metrics → git → effects/taint → health → search). First-time users now see what's happening instead of staring at a single progress bar.
• Troubleshooting docs anchors wired to error codes. The _DOC_LINKS map referenced fragments (#index-stale, #db-locked, etc.) that the docs page didn't actually have as id="..." anchors — agents clicking the URL landed on the page but scrolled nowhere. Eight ids added to the troubleshooting sections; map trimmed to fragments that actually exist. New tests/test_doc_link_anchors.py pins the invariant in CI.
• roam mcp-setup --write writes the config to disk. Previously the command only printed the JSON block; users had to copy-paste into the right file. The new flag resolves the platform's expected path (~/.codex/config.json, ./.vscode/mcp.json, etc.), merges the roam-code server entry with any existing mcpServers block (preserving other servers), and leaves a sibling .bak copy. Refuses to overwrite a corrupt JSON file.
• Skip git-history pass when HEAD unchanged. Saves 1-10s per warm roam index on big-history repos. Manifest's recorded HEAD compared against live git rev-parse HEAD.
• SQLite pragmas tuned. mmap_size=1GB (was 256MB), wal_autocheckpoint=10000 (was 1000), PRAGMA optimize on commit. Closes p50 query-latency drift after heavy index loads.
• FTS5 schema gains docstring column. roam retrieve and roam search-semantic now match against natural-language docstrings — previously the FTS5 BM25 path was blind to docstring text. Schema migration drops + recreates the table on first run after upgrade; build_fts_index repopulates. BM25 weight 4 (between qualified_name=5 and signature=2).
• Memoize _find_function_node in compute_and_store. Avoid the second AST descent per callable; ~10-15% indexing win on Python/TS files.

Agent / MCP DX

• roam_validate_plan MCP tool. Pre-apply validator for a multi-step change plan. Takes [{kind: "rename", symbol, new_name}, …] and returns a verdict (ok / needs-review / blocked) with per-operation blockers, warnings, and advice — symbol existence, name collisions, blast radius, target-file sanity, fitness violations. Cuts 4-call agent loops to one round-trip.
• roam_for_<situation> family. Four situation-keyed compounds: roam_for_new_feature (understand + search + context + complexity), roam_for_bug_fix (diagnose + tests + diff + context), roam_for_refactor (preflight + impact + complexity + clones), roam_for_security_review (taint + vuln + critique + adversarial). Each bundles 3-4 inspect calls into one round-trip.
• Reference-based handles for >50KB envelopes. Tools returning JSON > ROAM_MCP_HANDLE_KB (default 50) write the payload to .roam/responses/<sha16>.json and return a tiny envelope with {handle, byte_size, preview}. New roam_fetch_handle(handle) retrieves the full payload on demand. Content-addressed: identical responses share a handle. Stops a 70KB roam_understand envelope from blowing the agent's context budget when the agent only needs the summary.
• roam_ask MCP tool. Wraps the 24-recipe TF-IDF intent dispatcher so agents on MCP clients can dispatch a recipe in one call instead of falling back to Grep+Read. Added to _CORE_TOOLS.
• roam_session_metrics MCP tool. Local-only per-tool invocation telemetry (success / rate_limited / error counts). Helps answer "which tools are agents actually using?" without phoning home.
• agent_contract block on every JSON envelope. ~200-token derived block: {facts, risks, next_commands, confidence} so context-budget-tight agents can read just this and skip the full payload. Opt-out via ROAM_AGENT_CONTRACT_BLOCK=0.
• Soft contract enforcement on destructive tools. roam_mutate --apply checks session memory for a prior roam_simulate call against the same target and injects a contract_compliance block with actionable advice when the prerequisite was skipped. Soft-warn only, never refuses.
• Stale-index affordance on every read-only MCP tool. When the indexed DB is older than the manifest's recorded HEAD or older than the configured threshold, the response gets prefixed with INDEX STALE — call roam_reindex first. and a _meta.stale_index marker. Recovery commands (index, reindex, init, doctor, watch) skip the banner.
• summarize=True default for compound tools when ROAM_AI_ENABLED=1. roam_explore / roam_understand / roam_health now compress responses by default for users who've explicitly consented to MCP sampling. summarize=False forces opt-out per call; ROAM_AI_DISABLED=1 env var disables globally; the compliance preset always opts out (audit-trail evidence must be deterministic).
• @_tool decorator carries version="1.0.0". Surfaces in roam_catalog so agents can detect schema drift without re-enumerating every tool. Bump per-tool when the input/output schema changes.

Onboarding (roam init)

• No more unsolicited CI workflow file. roam init previously dropped .github/workflows/roam.yml into every repo on first command — the single biggest first-run trust-damage signal flagged by the audit. Now requires --with-ci=github to opt in; the existing roam ci-setup remains the canonical path for full multi-platform CI generation.
• Refuses outside a git repository (FILE_NOT_FOUND structured error). Prevents the spawn-.roam/-in-~/Downloads failure mode.
• Auto-writes a commented .roamignore template when absent. Every entry commented out so the user opts in to the patterns that apply.
• Compact welcome banner — 4 lines (stats + try-one + next + help), down from 20+ lines of agent-contract teaching at a moment when the user just wants "did it work?"

roam doctor

• Three-tier exit codes. 0 = clean; 1 = only advisory failures (cache age, cloud-sync, optional extras); 2 = at least one blocking failure. --strict promotes advisory to blocking for CI gates that require zero drift. Closes a real CI-noise gap where stale-snapshot warnings spuriously failed every roam-on-roam run.
• Issue-template-ready summary line. Roam X · Python Y · OS Z · M/N checks pass · A advisory · B blocking — paste-once into a bug report.
• Manifest drift hints expanded. git_dirty_hash drift surfaces as INFO; config_hash drift (roam config / .roamignore changed) as WARN.
• If this looks unexpected, run \roam doctor\`" hint embedded in 4 environmental error paths (DB open / schema corruption, config TOML parse failure, bundle verify failure, MCP missing-fastmcp error).

CLI surface

• roam --help collapsed to a 30-line Start-here panel. Previously a 154-line dump that scattered the 5 verbs across 38 entries in "Getting Started" and a 73-name "More Commands" comma-list. Long surface still available behind roam --help-all.

Security tier-2

• Predicate IRI moved to https://roam-code.com/spec/... — the owned, dereferenceable domain. Legacy roam-code.dev IRIs still verify via _LEGACY_PREDICATE_TYPES so old statements don't break.
• Strip username:token@ from git remote URL in cga subject. A repo cloned with https://x:ghp_PAT@github.com/... would otherwise leak the PAT verbatim into every signed CGA's subject.name.
• vuln_store ingest hardening. 50 MB size cap on scanner reports (refuses with structured error before loading); ESCAPE clause + escaped LIKE patterns so a hostile package_name like _ no longer match-explodes.
• taint_engine path_truncated flag. When the BFS exits via max_hops or the per-node 200-edge fan-out cap, the finding carries path_truncated=True so OpenVEX consumers map to under_investigation rather than vulnerable_code_not_in_execute_path.
• Hash-pinned mcp-server-card.json. New tests/test_mcp_server_card_hash.py fails CI on unintended drift — the agent-surface card is a real attack vector if tampered.
• CSP Reporting-Endpoints + report-to wired in _headers. Endpoint at /csp-report is provisional (CF Pages will 404 until a worker is wired up) but the directive is in place.
• dev/pin_github_actions.sh — one-shot script to pin every workflow action to a commit SHA via gh api. Defers actual pinning to a session with network; Dependabot's existing github-actions schedule maintains the pins thereafter.
• Keyless OIDC verification no longer hides under continue-on-error. The cga-attestation.yml keyless-oidc job now fails loud if the production-path emit/verify breaks. Sigstore outages are real but rare; masking flakes there hid regressions in our own emit/verify code path. The offline-key job remains the deterministic gate that runs without network.
• CycloneDX SBOM generation in publish.yml. Each PyPI release now ships an sbom/roam-code-<version>.cdx.json artifact built against a fresh venv with the just-built wheel installed (resolves to the runtime dependency closure, not the CI build env). Audit-trail product needs verifiable bills-of-materials; this is the foundation.

Architecture substrate

• @_tool carries version (see Agent / MCP DX above) — surfaces in roam_catalog.
• LanguageBridge.VERSION + LanguageExtractor.VERSION ABCs gain a 1.0.0 semver class attribute. Bump in subclasses when inference logic changes; downstream drift detection compares stamps.
• index_manifest step-completion record. Per-step success / failed:<ExceptionName> / skipped:<reason> status persisted as JSON in the manifest's notes field — roam doctor can later surface "your index is missing taint analysis because that step failed" instead of the generic stale-manifest signal.

Tests

• 9 new test files: test_db_user_version, test_doctor_hints_in_errors, test_extension_versioning, test_mcp_contract_enforcement, test_mcp_server_card_hash, test_mcp_tool_telemetry, test_mcp_tool_versioning, test_n1_fixes, test_surface_consistency.
• 12 test files extended with new cases for the corresponding source-side changes.
• Surface-consistency test locks the 8-way split-brain dict invariants today (every _COMMANDS entry has a _CATEGORIES entry; every _CORE_TOOLS member is a real @_tool declaration; every _DEPRECATED_COMMANDS.replacement resolves) before the larger Capability Registry rework lands.
• CountingConn-based query-count regressions for every N+1 fix — proves the batched form scales O(1) in input size.
• 2,612 tests pass on the touched-areas parallel sweep; 1,495 on the focused sequential sweep.

Conventions

• tests/conftest.py:make_src_project writes .gitignore with .roam/ so the dirty-tree refusal behaves the same in tests as in production.
• Categorise the formerly-orphan lsp command under "Refactoring" alongside stale-refs (the LSP server is the same engine surfaced over JSON-RPC).
• Help-text template ratchet (R8.G9). New tests/test_command_help_template.py pins a curated set of high-leverage commands (init, doctor, the five Start-here verbs, and daily staples — 12 today) to a consistent shape: 1-line summary, an Examples: block, and an inline "See also" cross-reference. The list grows incrementally; new commands are encouraged to match. Twelve commands polished in this round.

---

Earlier in the [Unreleased] window

• roam surface — canonical capability registry as JSON or text. Source of truth for commands, aliases, MCP tools, presets, categories, maturity. --filter stable|experimental|internal|deprecated and --category <name> flags. Used by docs generation, contract tests, release notes, and the marketing/landscape surfaces.
• roam explain-command <name> — per-command introspection: category, maturity, aliases, deprecation, MCP exposure, DB tables touched (best-effort source-grep), optional extras detected, stale-index sensitivity (high / medium / low / unknown).
• roam db-check — integrity sweep over the local index: orphan symbols, broken edges, duplicate file paths, missing FTS rows, invalid line spans, corrupt metrics, files with zero symbols. Exit code 5 on any high-severity finding when --ci is set.
• roam health --baseline <ref> — compares against a stored snapshot at the named ref (any git ref, last, or auto for default-branch). Reports delta (new / fixed / regressed findings) instead of an absolute set. Verdict reflects the delta — REVIEW on new high-severity, BAD on score regression, OK otherwise. Graceful DEGRADED when no snapshot exists.
• Index manifest table. New index_manifest schema records the indexer run: roam version, schema version, parser versions, grammar versions, config hash, git HEAD, dirty hash, enabled extras, index profile. roam doctor surfaces manifest age plus drift hints (parser-version drift, git-HEAD drift) so commands depending on graph accuracy can warn or refuse.
• roam doctor expansion. Four new checks: optional extras (onnxruntime / watchdog / fastmcp / scipy presence), cloud-sync detection (OneDrive / Dropbox / iCloud / Google Drive / Box / pCloud), cache permissions (.roam/ and .pytest_cache/ writable), and index manifest age + parser drift. 17 checks total.
• roam grep rewritten end-to-end with index-aware enrichment. Multi-pattern (-e repeatable, --patterns-from FILE), multi-glob (-g py -g md), -F/--fixed-string, -i/-w flags. Engine selection prefers ripgrep > git grep > indexed-file fallback (pin via ROAM_GREP_ENGINE). Bulk-fetch enclosing symbols once per file (interval index) — replaces the per-match SELECT N+1 path. New flags: --reachable-from <entry[,entry,...]> (forward BFS over the call graph), --unreachable (orphan / not-reachable filter), --co-occur (every -e pattern must hit the same enclosing symbol), --missing-pattern P (anti-correlation), --rank-by importance (PageRank-ordered output), --group-by symbol (collapse hits per fn/class), --blame (last author + date), --heat (file churn), plus opt-out --no-clones / --no-bridges. Each match carries clone-class siblings (from clone_pairs) and cross-language bridge links when applicable.
• roam refs-text <string>... — string-audit verdict shaped like safe-delete but for arbitrary literals (config keys, file paths, URLs, error messages). Groups references by surface (code / test / docs / config / dead), annotates reachability for code hits, emits per-string verdict (SAFE-TO-REMOVE / REVIEW / LOAD-BEARING). --reachable-from <entry> anchors reachability; --per-match-detail includes match lists in JSON.
• roam delete-check — gates the working diff on surviving references. Parses Python / JS / TS / Go signatures from deletion lines, searches the post-edit tree for surviving uses, classifies each deletion SAFE / LIKELY-SAFE / BREAK-RISK. --source {working,staged,pr,head}, --ci exits 5 on BREAK-RISK for CI gating. Pairs with the PR Replay narrative.
• roam history-grep <pattern> — git pickaxe wrapper (-S / -G) with -e repeatable, --since / --until, -p path filter, --polarity to mark each commit as introduced / removed / modified. Useful for postmortems and provenance audits when the trace is no longer in HEAD.