Chameleon

Added

• The review surfaces close the finding-fate feedback loop. pr-review Step 5b
  now records a refuted-dropped BLOCK/FIX as declined in the fate ledger
  (previously only surfaced/accepted and converted findings were recorded,
  leaving the pr-review per-lens precision denominator permanently empty), and
  all three review skills read get_finding_fate_stats back before their
  first refute_finding batch as an advisory, fail-open calibration:
  pr-review orders refuter sends within a severity by the repo's own per-lens
  refuter-survival history, receiving prioritizes a PUSH BACK on a comment
  category the repo historically applies, and deep-work prioritizes a decline
  that contradicts its lens's acceptance history — so the verdicts most likely
  to be wrong reach the independent refuter before the 4-call budget runs out.
  History never decides a fate, never gates a finding, and never appears as
  evidence.
• The three packaged agents (code-scout, web-researcher, pattern-reviewer)
  carry an explicit token-economy and parallel-dispatch contract: independent
  probes fire as one batch, the final message is the answer alone
  (pattern-reviewer: the JSON alone), and quoting is capped to claim-bearing
  lines.

Fixed

• pr-review Step 2.8 documented the Redux-slice co-change trigger as a
  lowercase substring match ("basename CONTAINS slice", citing
  sliceHelpers.ts as triggering). The engine matches the <alnum>Slice.ts(x)
  capital-S suffix only, so the skill's own example was false; the text now
  states the suffix rule with engine-true examples.
• Stale-vs-untrusted trust handling: a stale grant (revalidation opt-in)
  still serves profile content engine-side, but deep-work told the model to
  fall back to manual reading whenever trust_state was not trusted, and
  receiving lumped stale in with untrusted for plain-judgment fallback. Both
  now degrade only on untrusted/absent and carry a staleness caveat otherwise;
  deep-work also no longer claims the untrusted envelope returns the archetype
  name (it is withheld along with the content).
• pr-review Step 2b and the pattern-reviewer agent no longer suggest reading a
  "suggested fallback" archetype off the null-match get_pattern_context
  envelope — that envelope carries no such field (the receiving skill already
  stated this correctly).

Changed

• pr-review SKILL.md trimmed ~7% (61.1KB -> 57.0KB) by deduplicating the
  Integrity/Important/Honesty restatements against the authoritative
  references; deep-work's frontmatter description trimmed ~40%. No contract or
  pinned invariant changed (full unit suite green).