fix(review-pr): fail closed on incomplete outcomes

[Sayt-0]

Summary

• makes PR review outcomes fail closed instead of deriving success from an agent exit code or log text
• keeps every GitHub review submission on the COMMENT event and replaces approving language with the neutral 🟢 NO FINDINGS completion label
• validates and attributes each posted review with a trusted per-run marker, then propagates the classified outcome through the reusable workflow and check run
• preserves low-severity findings, bounds delegation retries, and updates incremental-review and rate-limit handling for the new review markers
• refreshes reviewer eval fixtures after model-backed canaries exposed stale or contradictory scenarios

Root cause

Several independent paths could present an incomplete run as successful:

• an exit code of zero without a posted review triggered a synthetic LGTM fallback
• low-severity findings were skipped during verification and could disappear before assessment
• incomplete or inconclusive agent output was not independently classified after posting
• the reusable workflow used job.status while continue-on-error masked review failures
• the posting template contained an approving default and route placeholders the model had to rewrite

Behavior

Scenario	Result
complete, conclusive, zero surviving findings	COMMENT with ### Assessment: 🟢 NO FINDINGS
any surviving finding, including low severity	COMMENT with NEEDS ATTENTION or CRITICAL
incomplete drafter coverage	COMMENT with Review incomplete; no incremental checkpoint
inconclusive verification	COMMENT with Verification inconclusive; no incremental checkpoint
no attributable review posted	non-approving incomplete notice and non-success workflow status
malformed, conflicting, approving, or non-comment review	fail closed as unverified

Safety and compatibility

• stages immutable repository, PR, and head-SHA routing in the trusted action
• creates and masks a cryptographically random per-run nonce
• validates the review body and comments array in bundled TypeScript before posting
• constructs the payload in a file before invoking gh api, with event: COMMENT hardcoded
• rejects active APPROVE, LGTM, and No issues found wording
• keeps historical ### Assessment: reviews compatible while rejecting the old synthetic LGTM fallback as an incremental checkpoint
• keeps structured-output schemas provider-compatible and free of JSON Schema pattern

Validation

Check	Result
Unit tests	40 files, 1413 tests passed
TypeScript typecheck	passed
Biome	passed
Build	passed
actionlint	passed
Shell integration tests	3 suites passed
YAML and eval JSON parsing	passed
Model-backed targeted evals on Docker Agent v1.121.0	4/4 evals, 24/24 relevance criteria
Model-backed clean-case repetition on v1.121.0	3/3 evals, 24/24 relevance criteria
Additional security, confidence, and scope canaries	3/3 evals, 20/20 relevance criteria
Fresh-eyes and specialist review findings	resolved

Model-backed eval outputs are intentionally kept outside the repository. No eval posted to a real pull request; GitHub posting behavior is exercised through the production shell harness with a mocked gh endpoint.