Chameleon

A hostile, exhaustive QA round across every supported language and framework
plus a code-quality audit of the plugin source, for a production-stability
release. 32 findings were confirmed and fixed: 1 correctness BLOCK, 22
gap/inconsistency/honesty fixes, and 9 slop cleanups. The language and
framework fixes were verified live by re-deriving the real Next.js, NestJS,
Rails, Django/DRF, Flask, and FastAPI repos; the deterministic fixes are pinned
red-green.

Fixed

• Cross-file existence check reported a false break (and could false-deny) on
  a name surviving only in a multi-line comment or template literal.
  _reference_present blanked a single recorded line in isolation, so a removed
  export whose name lingered inside a block comment or template that opened on an
  earlier line read as a live reference. It now blanks the whole file first, then
  indexes the line, matching the Ruby sibling. (hook_helper.py)

• Ruby module methods were invisible to the constant graph and call index. A
  def self.foo inside a module is invoked as Module.foo (the
  constant-receiver shape), but the AST dumper never pushed the module as a
  receiver identity, so Module.foo calls got no edge and the module's
  defined_in was empty. get_callers on such a method returned zero callers
  despite real call sites. The dumper now records the module as the enclosing
  identity for its singleton methods. (prism_dump.rb)

• NestJS DI call edges were dropped for a non-null-asserted or parenthesized
  receiver. this.foo!.m() and (this.foo).m() emitted no call site while
  this.foo.m() and this.foo?.m() did, so get_callers / get_blast_radius
  under-reported. The dumper now unwraps non-null assertions and parentheses on
  the receiver. (ts_dump.mjs)

• get_contract_breaks was blind to a required keyword-only argument in Python.
  A def f(*, x) addition breaks every caller (all pass by keyword) but the
  positional-only count stayed flat, so the break read as clean. Required
  keyword-only params now count as a narrowing for Python; Ruby stays unchanged.
  (signature_diff.py)

• Class-contract directive projected a minority cohort as an archetype-wide
  MUST. A base/decorator carried by a minority of an archetype's files (or a
  smaller cohort out-ranking the dominant one on richness) was rendered as an
  absolute "extends X; define Y" -- a DRF-view archetype was told to extend a CBV
  mixin 0 of its files use. The anchor now counts distinct files (matching its
  own threshold and the inheritance detector) and the dominant cohort always
  wins. (conventions.py)

• DRF authz advisory false-fired on a viewset guarded by a project decorator.
  A viewset with @allow_permission(...) on every action was flagged "declares
  none" because the authz vocabulary was Django built-ins only. A decorator whose
  name carries an authz token now satisfies the convention. (lint_engine.py)

• Framework detection mislabeled non-Django repos as Django. A bare
  manage.py forced 'django' before the dependency checks ran, so a Flask repo
  using the Flask-Script manage.py convention resolved to django. Dependency
  signal now decides first; manage.py counts only when its content is a real
  Django entrypoint. (orchestrator.py)

• get_archetype masked a corrupt profile as a clean no-match and had no trust
  gate. It collapsed a load failure to the same empty payload as a legitimate
  no-match (every sibling reports degraded), and served classification from an
  untrusted profile while all siblings refuse. It now reports degraded on a
  corrupt bundle, trust-gates like its siblings, and names a repo-arg mismatch.
  (tools.py)

• query_symbol_importers missed all broken importers of a deleted module. A
  deleted module's unreadable content returned found:false, lumping deletion in
  with oversized/unsafe. A deleted module now resolves to a closed empty export
  set so every still-referencing importer is reported, matching
  get_crossfile_context. (tools.py)

• The correctness judge's array parser could drop real findings to a decoy. It
  returned the FIRST object-containing array, so a decoy array in the model's
  reasoning shadowed the real findings emitted after it. It now keeps the last,
  matching its documented contract. (judge.py)

• A poisoned enforcement-state scalar bypassed fail-open and then crashed every
  save. Non-map scalar counters were read raw, so a tampered string survived
  load and crashed each subsequent (swallowed) save, silently losing enforcement
  accounting for the session. Scalars now coerce fail-open like the map
  counterpart. (enforcement.py)

• posttool-recorder armed the Stop backstop during a pause/disable window and
  wrote an empty exec-log row for every Edit/Write. It now honors suppression
  before recording a Bash-written file and only appends the exec log for the Bash
  tool. (hook_helper.py)

• The per-edit cross-file advisory surfaced only on the daemon path. The
  in-process lint fallback (daemon down, Bash writes, Stop re-lint) omitted the
  removed-export / importer advisory; it now runs there too, with the truncation
  flag threaded so a >100KB prefix does not false-flag. (hook_helper.py)

• The phantom-symbol check skipped multi-line parenthesized Python imports --
  the dominant multi-name style -- so a hallucinated name inside from m import (\n a,\n b\n) was never flagged. The import scan now spans the parenthesized
  body. (phantom_imports.py)

• get_rules, doctor, and the shadow/longitudinal windows carried honesty and
  robustness gaps: get_rules conflated an unresolvable repo with a healthy
  zero-rules repo; doctor false-cleaned on a bare-array or generation-mismatched
  core artifact; bootstrap_repo crashed on a non-string paths_glob; an untrusted
  repo's advisory surfaced secrets but not eval; the report window had no upper
  clamp. All now report or fail honestly. (tools.py, hook_helper.py,
  shadow_report.py)

Removed

• Dead code: refuter_available / refuter_unavailable_reason (superseded by
  refuter_cli_absent), the unused _line_at secret-scanner helper, and the
  never-read ClusteringResult.cluster_count / total_files_clustered
  properties. Stale doc/skill references (idiom_judge default, the recorder's
  drift-observation attribution, the engine version in architecture.md, the
  enforcement_artifact_unreadable status flag) were corrected, and the
  prose-rule miner now skips git-ignored ephemeral docs.