feat(search): passage-select + snippet-first for the answer path

[us]

Two gated levers for /v1/search LLM answer synthesis. Both default off, so
merging changes nothing in prod until the flags are flipped; off = byte-identical
to the prior answer path (A/B keep/revert).

Changes

• answer_bm25_select (new flag, default off): when a scraped source exceeds
  the per-source byte budget, keep its query-relevant passages (sentence-chunked +
  BM25-ranked, reusing the existing chunker/ranker) instead of a blind
  head-truncation that drops answers buried deep in a page. Two-pass with a
  partial-fill tail so it never feeds the model less than a head-truncation would
  (monotone on recall); non-adjacent passages are joined with a [...] gap marker.
• snippet_fallback (repurposed → snippet-first): prepend the SERP snippet to
  every answer source and block-guard the scraped body, so a fetched-but-blocked
  page (a "Wikimedia Error" / bot-wall shell) falls back to the clean snippet
  instead of grounding on error text. The legacy path (flag off) is byte-identical,
  so the multi-round scout still fires on such sources.
• Citation parsing splits on the last ===CITATIONS=== marker (rsplit) so a
  source that quotes the marker cannot truncate the answer or divert the parse.

Safety / rollout

• Both features gated, default off — no behavior change on merge.
• cargo fmt + cargo clippy --workspace -D warnings + cargo test --workspace
  all green; new unit tests cover passage-selection (deep-answer, budget-fill,
  oversized-blob, passthrough) and the block-shell detector.
• Opencore images build on release only, so this is not live on prod until the
  next release-please release; activating the levers is a separate config flip.