feat(code-review): synchronous GitHub-mode verifier dispatch + loud missing-verifier signal (FEA-3154)

[peterulsteen]

Summary

Applies the merged v3.6.0 reviewer-fleet hardening (#179 / FEA-3085) to the verifier fleet (stage_23_verify_findings). Bumps code-review 3.6.0 → 3.7.0. Implements PLN-1334 (approved).

stage_23 (skill code-review:verify-findings) spawned every verifier with run_in_background: true + blocking TaskOutput in both modes — the exact pattern v3.6.0 proved unsafe under headless claude -p, where the orchestrator can end its turn before issuing the blocking collection, so the process exits (terminal_reason: "completed") and verifier verdicts never land. This is worse for verifiers than reviewers: stage_23 is on_failure: continue, so a missing output degrades silently to pending_verification[], which _compute_canonical_verdict does not read — a BLOCKING finding whose verifier died could ship APPROVED.

Changes:

• skills/verify-findings/SKILL.md — MODE branch: github mode dispatches verifiers synchronously one-at-a-time (run_in_background: false, wait for each Task response before the next), local mode keeps parallel background + blocking TaskOutput; adds a headless-mode warning. Synchronous dispatch keeps the headless turn alive by construction (each Task is itself the pending call).
• commands/start.md — stage_23 stage-note sync clause + a Verifier Fleet walker guard mirroring the stage_20 guard (bans watcher files / sleep loops / polling loops / turn-ending waits).
• cmd_verify_consolidate — defense-in-depth: in github mode, when ≥1 BLOCKING/HIGH finding has no verifier output, emit one aggregate coverage:verifier-missing-output coverage-gap finding (severity: HIGH, required: false) → _compute_canonical_verdict escalates to NEEDS_ATTENTION (mirrors the existing "verifier uncertain → human triages, not silent approval" semantics). New _make_unverified_findings_gap helper; new optional --mode on verify-consolidate, wired at stage_24a_verify_consolidate (same {mode} token finalize-result already uses). Gap id continues past any pre-existing coverage_gaps.json entries (no collision with stage_20b's coverage-verifier_f*).
• Tests — TestFEA3154VerifierDispatchContracts prose-contract class (mirrors FEA-2162) + behavioral cmd_verify_consolidate/cmd_finalize_result cases (github HIGH emits gap + end-to-end finalize asserts validation_errors == [] and NEEDS_ATTENTION; local/MEDIUM/deferred-budget negatives; count aggregation; index continuation). Regenerated the 3 declarative-config snapshots (only the --mode addition).

Net-new github-synchronous verifier behavior — not a revert (pre-v3.5.0 verifiers were background in both modes).

Reviewed pre-PR by an adversarial Codex deep review (multi-worker: verdict/schema/prose/id-index lanes) and a /code-review --depth deep pass — both APPROVED with no BLOCKING/HIGH/MEDIUM findings. Their two test-depth suggestions (count aggregation, index continuation) are included.

CHANGELOG note: the v3.7.0 entry was hand-reconciled to the /update-documentation generator's format (version heading matches plugin.json, newest-first, single root CHANGELOG.md) because this branch was authored from a headless session that can't invoke the plugin slash command — same approach accepted on #179. Happy to regenerate via /update-documentation before merge if preferred.

Feature Flags

N/A — plugin behavior change, no runtime flag. --mode on verify-consolidate is optional and back-compatible (omitted → mode=None → no signal, prior behavior).

Breaking changes

None. Omitting --mode preserves prior behavior. Trade-offs: github verifier stage wall-clock becomes sum-of-verifiers (was slowest-of, but that path was unreliable); a missing github verifier for a BLOCKING/HIGH finding now escalates the verdict to NEEDS_ATTENTION instead of a possible silent APPROVED.

Test plan

• uv run --frozen --group dev pytest plugins/code-review/tools/python/test_code_review_helpers.py → 1177 passed
• uv run ruff check . → clean
• uv run pyright → 0 errors, 0 warnings
• Regenerated config snapshots verified diff = only the --mode addition
• Codex deep review + /code-review --depth deep subagent → both APPROVED

🤖 Generated with Claude Code

[mikeangstadt]

Solid. The synchronous-primary + coverage-gap-defense split is the right shape, and I traced the loud path end to end: gap lands in coverage_gaps.json, finalize reads it, Rule 3 escalates the HIGH gap to NEEDS_ATTENTION so a dead-verifier BLOCKING/HIGH can't ship APPROVED. Contracts are all in sync (skill, start.md, cli.json, stages.json wired at stage_24a, snapshots, version bump). One tiny doc nit inline, not blocking.