Docs/condense oversized docstrings

[brunovicco]

Seventeen modules carried docstrings that had grown into design essays -run.py alone opened with 152 lines before the first import. They restated decisions the ADRs already own, recorded implementation history, and
documented parameters by paraphrasing their names.

Docstring lines: 1,091 → 676 (−38%). Nothing else changed.

Kept: observable contracts, Raises, and invariants a caller has to know - fail-closed preflight, the judge factory's thread-safety requirement, the consecutive-error circuit breaker, unanswerable questions never dropped from
coverage, the HNSW ≤2000-dimension constraint behind the Matryoshka request, the checksums-before-manifest ordering that keeps a crashed run from looking complete, and L2 normalization matching pgvector's cosine operator.

Removed: re-explanations of ADR decisions (one reference replaces the paragraph), rejected-alternative narratives, and Args: blocks that only restated parameter names.

chunking/legal_parser.py was deliberately left alone: its notes on legal structure parsing describe behavior no ADR covers.

Also corrects a stale claim in ragas_judge.py, which said the golden set has no reference answers as the reason Factual Correctness is unimplemented - v0.2 has one for all 230 questions. Now scoped as "not wired in this
increment", which makes the metric a live candidate for the next one.

Verified: ASTs with docstrings stripped are identical to the originals, file by file - no code changed. ruff check --select D,E501 and ruff format --check pass with the repo's own config.