Chameleon

Added

• Turn-end VERIFY stage: the automatic review now refutes its own findings
  before you see them. The turn-end pipeline ran SCOPE -> EVIDENCE -> ATTACK
  -> REPORT with no VERIFY: the correctness judge's findings surfaced
  unchallenged, while the fully-built independent refuter (refuter.py) was
  wired only to the interactive pr-review / receiving skills — the plugin's
  most-used review surface had no independent verification at all. A new
  stop_verify.py module closes the gap on all three turn-end paths: the
  default multi-lens pass (lone-correctness-lens findings only; cross-lens
  agreement is already an independent signal and duplication findings' evidence
  spans two locations a one-file excerpt cannot show, so those are exempt), the
  single-lens correctness gate, and the async-detached judge child. A refuted
  finding is dropped, a confirmed one is tagged [confirmed], the rest surface
  labeled unverified, and the advisory carries a grounding banner
  (N refuted and dropped, M confirmed). Strictly raise-precision by contract:
  only an actively-refuted finding is ever removed; any failure (no budget, CLI
  absent, spawn error) passes everything through unverified; and a finding
  whose excerpt cannot be safely fetched is never spawned, because the refuter
  prompt commands refutation on cannot-tell and a zero-evidence spawn would
  kill anchorless findings on engine-side data absence. Budget-honest under the
  55s Stop wrapper: refuter spawns are retry-free (run_one(retry=False), new
  parameter) so a slot costs exactly one timeout window, and the sync path
  anchors its remaining-budget math to process start — the same clock as the
  wrapper's SIGKILL — spawning only when a full window fits. Excerpts are
  containment-checked through safe_read_text (the finding's file field is
  model output inlined into a prompt; an absolute or ..-escaping path reads
  nothing). Raw findings still shadow-log pre-VERIFY so precision sampling
  keeps seeing the refuted rows; only survivors persist to the finding ledger.
  Default-ON with the CHAMELEON_STOP_VERIFY=0 kill switch; no repo-code
  execution and no network (the refuter is a hardened no-tools claude -p).
  Verified live end-to-end: a deliberately false claim came back refuted and
  dropped, a real defect came back confirmed and kept, ~8-10s per spawn.

• The five-phase review discipline is now explicit in /chameleon-pr-review.
  The skill already ran all five phases; a new "The five-phase discipline"
  section names them (SCOPE -> EVIDENCE -> ATTACK -> VERIFY -> REPORT) and maps
  each existing step to its phase, so the skill and the engine's turn-end
  pipeline now describe the same contract in the same terms.