cc-triage below-floor skip — playground (#1739)

Smallest-diff fix: cc-triage.mjs now reads shared/cc-support.json at startup and skips the LLM call for any version below supported_floor. Marks the entry with below_floor: true; the workflow's existing jq 'select(.gap_score >= 10)' filter naturally skips entries with empty features[], so issue filing already does the right thing.

Real-run evidence job timeout

cc-watch run 25632103449 (2026-05-10) attempted to triage every unsnapshotted version — W1b's snapshot-presence filter exposed 100+ ancient versions. The job hit GH Actions' cancellation timeout before finishing.

cc-triage: extracting features from 2.0.59...
cc-triage: extracting features from 2.0.60...
cc-triage: extracting features from 2.0.61...
[... 100+ iterations ...]
##[error]The operation was canceled.

Try it interactive

supported_floor candidate version

Implements the same compareSemver used in cc-release-watch.mjs — numeric-tuple compare, no prerelease handling needed (CC versions are X.Y.Z).

Decision matrix

Condition	Action	features[]	parse_failed	below_floor
version < supported_floor	SKIP LLM	`[]`	(unset)	`true`
version >= supported_floor, snapshot present	CALL LLM	extracted	`false`	(absent)
version >= floor, LLM fails twice	CALL LLM	`[]`	`true`	(absent)
cc-support.json missing/malformed	CALL LLM (legacy)	extracted	`false`	(absent)

Why a sentinel field instead of dropping the entry

Two reasons:

cc-release-watch emits every unsnapshotted version into the gaps file as a recovery hatch (W1b). Dropping below-floor entries from cc-triage's output would re-introduce the silent-loss bug W1b fixed.
parse_failed means "LLM tried and failed" — that's a retryable state via --retry-failed. Below-floor is a policy decision, not a failure. Conflating them would cause --retry-failed to repeatedly burn LLM calls on ancient versions.

Files changed

Path	Change
scripts/cc-triage.mjs	Add `compareSemver` + `readSupportedFloor` helpers; per-version loop sets `below_floor: true` and skips `callClaude` for sub-floor entries
tests/integration/test-cc-triage.sh	New Test 8 covering W1h: below-floor entry skipped (no LLM call, no `parse_failed`), above-floor entry processed normally, no flag leak between entries

Verification