Chameleon

Per-edit injection enrichment from EFFECTIVENESS-REVIEW-2026-06-22, a measured
A/B study showing chameleon lifts first-try convention conformance from 1/8 to
8/8 on house rules a model cannot guess. This release ships the proven-direction,
low-risk recommendations; the cross-file collaborator-signature work (R1) is
grounded and planned but gated on an effectiveness A/B before it ships, because
its own cited research ("more retrieval can hurt") means it must be measured, not
assumed.

Changed

• The witness imperative is now calibrated by match quality. A strong
  structural match (exact/ast) tells the model to mirror the canonical witness
  closely; a weak match downgrades it to a loose reference and points at the team
  idioms as the repo truth regardless of file shape. Previously match_quality was
  printed in the header but never changed the instruction. The directive is
  emitted outside the untrusted spotlight region (it is chameleon's instruction
  about the data, not data to imitate).
• Team idioms in the per-edit block are capped and deduped against the
  witness. Idioms are bounded to the same 1500-char budget the PostToolUse path
  already used (the PreToolUse block had no cap), and any idiom line the canonical
  witness already demonstrates verbatim is dropped: repeating what the witness
  shows is noise that dilutes the model's attention on the whole block. Bounded
  substring dedup, no hot-path I/O added; skipped when there is no witness.

Block ordering was already match-quality-driven (the higher-signal section leads
for primacy), so the calibrated imperative is the ordering signal the review
asked for; no separate reorder mechanism was added.

Added

• Experimental: nearby collaborator signatures in the per-edit block
  (CHAMELEON_NEARBY_SIGNATURES=1, default OFF). The cross-file lever the
  review measured (it flipped a cross-file call from wrong to correct): instead
  of just sibling filenames, inject the real callable signatures of source files
  in the edited file's directory, read from the precomputed
  symbol_signatures.json (no live parse, mtime-cached, so it stays on the
  <100ms hot path). Off by default and env-gated pending an effectiveness A/B,
  because the review's own research ("more retrieval can hurt") means the lift
  must be measured before it ships on. Bounded (5 files, 8 signatures, 700
  chars), fails open, rides the existing sanitize + spotlight path. The default
  path short-circuits on the env check, so it adds zero cost when off.

Drift negative-examples (R5) and the test-shape witness / per-edit re-ranking
(R6/R7) remain deferred pending the same A/B.