test696 — authenticated human short replies are not silently dropped (#696)
Date: 2026-08-13
Environment: Docker, oven/bun:1.3.14
Scope: agent-node outbound low-value filtering and authenticated Dashboard provenance wiring

GREEN
Command:
  sg docker -c 'docker build --build-arg SOURCE_COMMIT=$(git rev-parse HEAD) -f tests/test696-human-low-value-reply/Dockerfile -t anet-test696:local . && docker run --rm anet-test696:local'
Result:
  - routing and authenticated provenance tests: 27 pass, 0 fail
  - production CLI provenance wiring: PASS
  - production bundle build: PASS
  - container exit: 0

WITNESSED RED
Pre-fix result:
  - "在线吗?" -> "在线。" from an authenticated Dashboard task: FAIL
  - received { shouldDeliver: false }, expected { shouldDeliver: true }
  - agent-to-agent task reply "收到": PASS (remained filtered)
  - targeted suite: 13 pass, 1 fail; exit 1

BROAD NON-GATE ATTEMPT
  A generic `bun test src` run in the narrow test696 image produced 1198 pass,
  25 fail, and 2 setup errors. It is not a valid full-suite gate: that image
  intentionally omits the agent-network tree, crontab, Grok fixtures/binary,
  OpenCode runtime, and root ownership required by those suites. No failing
  assertion involved the new reply-provenance behavior. The dedicated Docker
  gate and production bundle build are the acceptance evidence for this PR.

GREEN INVARIANTS
  - Authenticated Dashboard task/broadcast short replies bypass low-value filtering.
  - Agent-to-agent task replies do not bypass filtering merely because type=task.
  - A provenance flag paired with a non-task type does not bypass filtering.
  - Failed replies and Dashboard native slash notices retain their prior delivery behavior.

BOUNDARIES
  - No production DB, node configuration, or process was touched.
  - No live Hub request was required; the gate uses the Hub-stamped provenance predicate tests.
  - This test does not validate a production deployment or npm publication.
