Readiness fixes: fail-closed gaps, claims true-up, repo hygiene (audit 2026-08-24)

[bGOATnote]

Implements the P1 findings and P2 polish from the 2026-08-24 public-repo readiness audit.

P1

• P1-1 .gitignore now excludes .env, .env.* (keeping .env.example), .venv/, *.pem, *.key, .coverage — the quickstart creates .env at the repo root and nothing guarded it. Folds in the .gitignore lines PR Test coverage analysis + gitignore .venv/.coverage #1 proposed.
• P1-2 agent_loop treats stop_reason="tool_use" on the final message (SDK runner exhausted max_iterations with the model still requesting tools) as LLMFailure; run_compliance_review raises LLMFailure("...agent never called report_findings") when the summary holder is empty. A compliance audit can no longer render as a clean pass with zero findings.
• P1-3 structured() wraps json.loads so malformed JSON raises LLMFailure (the documented contract) instead of escaping as a raw JSONDecodeError → HTTP 500.
• P1-4 README metrics line now names the run explicitly (single run, 2026-07-18, on 12 of the current 13 goldset cases; rerun pending) and the test count matches the collected suite (297). Deferred to owner: rerun python -m halo.mci.demo on the 13-case goldset and commit evals/goldset_<date>.json (needs a live key; not run here).
• P1-5 docs/ECOSYSTEM.md CME box no longer says "MD-signed": hand-built; physician sign-off pending (all four cards review.status="draft").
• P1-6 synthetic-ambient-fhir-25/NOTICE states the dataset was provided by Abridge for hackathon use and that redistribution terms are unconfirmed. Deferred to owner: confirm terms with Abridge; the data itself is untouched.

P2

• SECURITY.md (disclosure contact b@thegoatnote.com) + short CONTRIBUTING.md.
• CI actions pinned to commit SHAs (checkout v4.4.0, setup-python v5.6.0).
• anthropic>=1.0,<2 (code depends on 1.x beta surfaces).
• The 5 mypy-strict errors in src/halo/mci fixed — mypy strict now clean on all 36 source files.
• Scenario-count consistency: 6 scripted scenarios (README + demo.py).
• STATUS.md archived to docs/HACKATHON_LOG.md; README/CLAUDE.md references updated.
• HandoffRequest.incident_date defaults to today (UTC), not the hard-coded hackathon date.
• edu/drill.py / edu/lookup.py log at warning level before failing closed.

Verification

• 297/297 tests pass offline (was 291; +6 new fail-closed tests)
• ruff check + ruff format --check clean
• mypy strict: no issues in 36 source files
• New .gitignore verified with git check-ignore (.env.example still tracked)

🤖 Generated with Claude Code

https://claude.ai/code/session_01142EwVwkXFRGnYu7YgHSYL