Run the regression suite automatically (CI + SessionStart hook)

[brianleach]

What

Makes the stdlib regression suite (tests/test_fixtures.py, added in #2) run automatically instead of only by hand, so the classifier→plumbing contract is actually enforced.

Changes

• .github/workflows/tests.yml — runs the suite on every push to main and every PR. No install step (the project is standard-library only); it just checks out, sets up Python, and runs the fixtures.
• .claude/hooks/session-start.sh + .claude/settings.json — a SessionStart hook for Claude Code on the web. Since there are no dependencies to install, instead of an install step it runs the suite as a fast, side-effect-free smoke check, so a web session starts knowing the fixtures are green.

Validation

The hook was run exactly as it will run in a web session:

$ CLAUDE_CODE_REMOTE=true CLAUDE_PROJECT_DIR=. ./.claude/hooks/session-start.sh
ok    real_inbox_results.json: exit=20 high=1 maybe=0
ok    sample_results.json: exit=20 high=1 maybe=1
ok    maybe_only_results.json: exit=0 high=0 maybe=1
ok    all_marketing_results.json: exit=0 high=0 maybe=0
ok    side-effects: --dry-run created no state/

All 4 fixture(s) passed.

No linter is configured in this project (stdlib only, no ruff/flake8/black), so the test run is the validation.

The hook runs synchronously — it guarantees the suite has passed before the session starts (it takes well under a second). Easy to switch to async if faster startup is preferred. Once this merges to the default branch, all future web sessions will use it.

🤖 Generated with Claude Code

---

Generated by Claude Code