docs: add per-package module maps and fix three stale facts

[sturlese]

Summary

docs/ explains how the system works; what was missing was the code map — where things live, what to reuse, what not to do — sitting next to the code. This adds an index.md to every package (nine: the pipeline stack + its five packages, answer, evals, benchmark), indexed from docs/README.md. Written by the documentator agent; every factual claim in this PR was verified against the code before merging.

Changes

• Nine module maps, each covering: purpose · entry points · module map · use these (the shared primitives, so the next contributor doesn't write a second one) · anti-patterns with the ADR that motivates them · data contracts · tests · common-tasks table.
• Three factual drifts between docs and code, fixed (the repo's own rule: when they disagree, the code wins — then fix the doc):
  • README claimed ~390 tests; the suite is 421 (fetch 29, clean 248, corpus 48, graph 38, slack 13, answer 42, benchmark 3) → ~420.
  • docs/architecture.md's stage table omitted slack, though the README and docs/pipeline/slack.md both cover it.
  • The README's "5 minutes" table displayed paths like clean/src/verify.py that don't exist — the hrefs were correct, the visible text wasn't, so anyone copying them got nothing.
• One non-obvious behavior now recorded: the golden scorecard is 24 metrics offline but 23 with a live model — eval_clean_and_trust swaps its two seeded-defect metrics for one when CLEAN_LLM isn't fake-flawed, since there's nothing seeded to catch. Whoever first evals a real model would otherwise think a metric vanished.

Testing

Docs only — no code touched. Claims spot-verified against the source: the make eval → run-evals.sh → evals/.venv chain, service.refresh() being called per ask_brain/search_brain request, index.superseded_paths existing, and the per-package test counts (run locally, all suites green).

Notes

Deliberately not done, flagged for your call:

• A root CLAUDE.md — the biggest remaining gap for agent onboarding (package layout, make targets, per-package venvs, the bare-pytest pythonpath gotcha, single-writer-per-artifact, never read .env). It's agent configuration rather than project documentation, so it shouldn't land without you deciding you want it.
• make test-pkg PKG=... shortcut target — a Makefile change, not a docs one.
• Documenting the facts.db schema in one place — currently implicit in clean/factstore.py and re-read by answer/metrics.py, with parity enforced only by eval_contract_parity.