Fix decision/error extraction gaps found by the external benchmark

[Shweta-Mishra-ai]

What this PR does

An independent 100-session, 8-method benchmark (Shweta-Mishra-ai/tokenmizer-research) scored this extractor's two weakest categories — decisions (50% F1) and errors (36% F1) — well behind two comparison methods (65% and 66%). This PR fixes the specific gaps that benchmark identified, relocated to patterns.py since an earlier attempt at this fix (superseded, never merged).

• Adds _is_question_context, wired into all four decision passes: a question weighing options (e.g. "Should we go with Postgres or Redis?") was previously extracted as a decision made.
• Broadens the decision trigger-verb vocabulary ("leaning toward", "recommend(ed/s)") and adds a handful of verified-safe technology names to the closed whitelist (sqlc, dbt, nats, pnpm, uv, ruff, kong, airflow).
• Expands the error-symptom vocabulary (nil pointer dereference, gc pressure, poison message, consumer lag, schema drift, partition skew, goroutine leak, connection churn, thundering herd) and widens the trailing-context capture after a symptom.
• Rejected and reverted two more aggressive fixes tried during this work: a whitelist-free CamelCase/kebab-case/ALLCAPS tech-name fallback (dropped decision precision 90%→81% on the eval corpus), and adding vite to the whitelist on its own (dropped precision to 88% via a false positive on "Completed: project scaffold with Vite"). Both are documented as negative results in the CHANGELOG.
• Version bumped 0.5.3 → 0.5.4 with a full CHANGELOG entry.

Type

• Bug fix
• Extraction improvement (graph_memory/)
• New provider
• Performance
• Documentation

Tests

• pytest tests/ -v passes — 644/647 (3 known-unrelated Gemini/cffi native-library failures, pre-existing and unrelated to this change)
• ruff check tokenmizer/ clean
• Memory accuracy test added/updated — python -m benchmarks.eval confirms no regression: decisions steady at 90% precision / 95% recall, errors 93%/96% (both matching or exceeding the pre-fix baseline). The product's own 14-session corpus doesn't contain the specific gaps this fix targets, so it verifies no regression rather than the improvement itself — that signal comes from re-running the external n=100 benchmark, where TokenMizer's macro F1 rose 57%→60% and decisions/errors F1 rose 50%→59% / 36%→44%.

Checklist

• No raw dicts crossing layer boundaries (use DTOs)
• No os.getenv() outside config/settings.py
• External imports are lazy (inside functions, with try/except ImportError)

---

Generated by Claude Code