v0.29.0 — BM25 rescue lane was silenced, not down-weighted

Changed

• The BM25 lane was silenced, not down-weighted, and is now audible.
  FTS_LANE_WEIGHT moves 0.3 -> 0.9. At 0.3 the lane could not surface a single
  row the vector lane had not already found — arithmetic, not corpus luck: with
  k=60 and fetch_limit=3x every vector-only RRF score sits in [1/90, 1/61]
  while an FTS-only hit at BM25 rank 1 scores w/61, so reaching position p
  needs w > (k+1)/(k+p) (0.87 for position 10, 0.97 for the top 3). At 0.3 an
  FTS-only row scored 0.0049 against a worst-case vector 0.0111 and could only
  reorder rows both lanes already shared.
  • Measured on 300 identifier queries over a real note corpus:
    Hit@1 35.3% -> 49.0%, Hit@5 46.3% -> 64.7%, MRR 0.402 -> 0.561.
  • Measured on 179 doc-block queries: unchanged (Hit@1 95.5%, MRR 0.973).
    0.9 is Pareto-dominant on both measured corpora — no regression found.
  • No migration, no re-embed, no schema change: applied at query time, so
    existing installs pick this up without reindexing. Result ORDER will change,
    which is the point.
  • Known limit, surfaced by an independent review before release: at w=0.9
    an FTS-only row can reach rank 8 at best, so a caller requesting fewer than
    8 results still won't see a BM25-only hit. This is strictly non-regressive
    versus 0.3 (which never surfaced one at any rank) — flagged as a target for
    a future confidence-scaled weight, not attempted in this release.

Added

• scripts/build_paraphrase_fixture.py, scripts/sweep_ranking_constant.py,
  scripts/probe_rerank_headroom.py — the unbiased benchmark fixtures and
  sweep tooling used to calibrate the constants below.

Fixed

• Constant documentation now records what was measured. VECTOR_GATE_THRESHOLD
  stays 0.82 (validated). RECENCY_BONUS_MAX stays 0.05 (peak of a unimodal
  curve; upside is noise, downside is real). PROVENANCE_HUMAN_BONUS is
  documented as UNCALIBRATABLE (898/905 active notes are agent-provenance).
  RRF k is documented as not a useful dial in this range.