Chameleon

Stop-hook correctness pass: kill a recurring "stale index" false positive across
all three languages, fix a SubagentStop that stole the parent turn's review, and
repair the turn-end reviewer health signal in both review paths. Every fix was
verified end to end through the real Stop hook against real profiled repos.

Fixed

• The turn-end cross-file existence advisory falsely reported "you removed X,
  still imported by ..." after a move-and-reimport refactor (the recurring
  "stale index" complaint). The reverse index is a bootstrap snapshot; when a
  symbol was MOVED to a new module and its call sites were repointed in the same
  session, the index still attributed the import to the OLD module and the
  advisory only checked whether the bare name still appeared in the importer (it
  does, now imported from the new module), so the phantom finding re-fired on
  every Stop until the next /chameleon-refresh. The advisory now resolves each
  importer's CURRENT import of the name with the same per-build specifier resolver
  the reverse index used and suppresses the finding when the name is imported but
  no longer from the edited module. A genuinely dangling import still fires; a
  parse miss falls back to the prior bareword behavior so a real break is never
  hidden. TypeScript (named imports, import {A as B}, re-exports), Python
  (from x import y, relative and absolute, single-line AND multi-line
  parenthesized via an ast parse), and Ruby (a class/module moved to a new
  file edited the same turn is no longer reported broken, since Ruby resolves the
  constant globally).
• A SubagentStop fired the once-per-session idiom-review block and stole the
  parent turn's review. The reflexive idiom/principle gate was the only blocking
  gate in the Stop pipeline not guarded by is_subagent: a subagent both got a
  spurious block on its narrow task AND burned the once-per-session marker, so the
  real top-level Stop then short-circuited and the turn-end self-review the
  enforcement exists to force was silently skipped. The gate is now top-level Stop
  only, matching every other top-level-only gate (multi-lens, duplication,
  scope-drift, attestation).
• The turn-end reviewer health signal was broken in both review paths. Under
  the default multi-lens path the correctness lens ran with no event sink, so a
  silently-dead reviewer (broken auth, missing binary) emitted no degraded-spawn
  event and the SessionStart health banner could never fire. Under the opt-out
  separate-gate path the reverse happened: the per-spawn grounding events
  (imported-definition and transitive-caller context availability) were
  mis-counted as degradations, so a perfectly healthy reviewer recorded a
  degraded spawn, raised a false "reviewer failed to spawn" banner next session,
  AND flipped the spawn-failed flag so the duplication gate stopped deferring and
  fired a SECOND reviewer model in the same Stop (toward the 55s wall-clock cap).
  Both paths now record the grounding families as their own check events and the
  degraded tally only on a genuine spawn failure, through one shared classifier so
  the two paths cannot drift again.
• Enforcement mode off emitted shadow-mode would_block telemetry. The Stop
  backstop's shadow branch covered both shadow and off, so an
  advisory-only off repo recorded would-block rows (misleading on a repo where
  enforcement is turned off). The telemetry is now gated to shadow; off stays
  fully silent, matching the idiom gate's handling. Neither mode blocks.