v2.8.1

Fixed

• Recall no longer serves confident noise on topics the store barely
  covers, and its reason string cannot lie. When the candidate head's
  cosine spread collapses (nothing to distinguish), the degree term is
  dampened proportionally instead of silently deciding the ranking, and
  the response carries a flat_cosine hint — hints are now actually
  serialized over the recall RPC (they were declared in the tool schema
  but never sent). The aaak overlap matches cues against content fields
  only — entity anchors and doc names — never machine tokens or
  bookkeeping tag values, so a cue word like "user" cannot bias every
  record of one role; entity anchors stay dormant until the capture path
  writes them. reason is assembled during scoring with weighted terms
  and the full multiplier trail (trigram, fts, lex fusion, tier,
  temporal, profile gain, stability, valence, corrector anchoring), so
  the printed arithmetic reconciles with the served score.
• Served recall ordering is now deterministic. Equal-scored hits used to
  inherit candidate scan order (Python's sort is stable), so the same cue
  could return a different ordering across calls; the supersede-cap window
  had the same tie sensitivity, which could shift served scores. Every
  serve path — pipeline, bank fallback, and the daemon's authority merge —
  now sorts through one shared helper that breaks ties on record_id, and
  a guard test keeps score-only sorts out of those files. Reported by
  @danielhertz1999-bit.
• FHRR bundle golden parity now states its platform tolerance instead of
  encoding one machine's libm rounding: components whose mean angle sits
  on a quantisation boundary (byte inputs make that exact) may differ by
  exactly one step, components with a near-zero resultant are skipped
  (atan2 is undefined there), everything else stays byte-exact — so
  formula and rounding-mode regressions still fail. Goldens unchanged.
  Measured and reported by
  @danielhertz1999-bit.
• A read-only reader holding a stale snapshot now raises the typed fence
  error the caller expects, rather than the older integrity shape. The
  crash-recovery test covering it had been red since the fence was typed.

Changed

• The Rust workspace is rustfmt clean, so formatting can be enforced in
  CI from here on.
• Two test-suite races removed: the multiprocess WAL test no longer
  depends on process interleaving, and detection-arena isolation is
  measured with a fresh process per arm.