Let a review comment declare several BCQuality articles

[Wenjie Fan (gggdttt)]

Problem

ReviewComment.article was singular, so an expected comment could be attributed to exactly one BCQuality article. That does not match how the gold answers are actually written: a number of them pair a positive requirement with an explicit false-positive boundary in one comment body, and those are two separate articles in BCQuality.

synthetic__error-handling-errorinfo-actionable-boundary-01
  "Raise ErrorInfo with RecordId = SalesLine.RecordId, PageNo set to the Sales Lines page,
   and a navigation action ... Do not require a Fix-it for Synthetic Save Target: that
   validation occurs on the selectable field itself ..."

synthetic__events-ishandled-reset-boundary-01
  "Start this raise from false and let subscribers opt in. Do not flag ValidateQuantity's
   fresh local or ValidateQuantityAndLocation's second raise ..."

This is deliberate, not incidental — the instance ids literally carry -boundary-. As the article corpus grows, one finding mapping to several articles becomes more common, not less.

Routing the second article through metadata.articles looked like an escape hatch, but it is a dead end. It holds only while no other comment in the entry owns that article:

case 1 (one comment needs A+B):     OK    declared=[A, B]
case 2 (c1 needs A+B, c2 needs B):  FAIL  article(s) [B] declared both per-comment
                                          and in metadata.articles

Case 2 is ordinary — two comments in one entry citing the same article — so the workaround breaks in normal use.

Change

• ReviewComment.article: ArticleId | None -> ReviewComment.articles: list[ArticleId].
• Migrated the 17 existing annotations in dataset/codereview.jsonl to the list form. No compatibility alias is kept; the field is days old and the dataset is fully in-repo, so carrying a shim would just be debt.
• Set extra="forbid" on ReviewComment. Pydantic ignores unknown keys by default, so without this a leftover "article": "..." would be silently dropped — a stale annotation would become invisible coverage loss instead of an error. Verified no comment payload in the dataset carries any other extra key, and both programmatic construction sites (collection/collect_codereview.py, evaluate/review_parsing.py) pass explicit keyword arguments.
• declared_articles() and the disjointness validator now flatten per-comment lists; metadata.articles keeps its existing false-positive-guard role and the disjointness rule is unchanged.
• Refreshed the notebook outputs, whose committed run still reported 130 entries (it predates Add second PR reviewer self-improvement gold batch #803). Now 9 articles covered across 17/144 annotated entries; 241 of 250 articles have zero coverage.

Validation

• Migration audited structurally: parsed all 144 entries before and after, compared every field except the article keys — zero drift; 17 annotations moved across 13 entries, and the article set per entry is unchanged.
• New tests cover a comment declaring several articles, and that a superseded singular article key is now rejected rather than ignored.
• ruff format, ruff check clean; pre-commit run --files notebooks/code-review-coverage.ipynb fully green; ty check reports only the pre-existing src/bcbench/redteam.py:150 unused-ignore warning.
• pytest -q: 740 passed, 2 skipped.