Chameleon

Remediation of the QA-30 full-surface campaign: a complete plugin inventory plus a five-lane hostile QA pass (failure/recovery, language depth, enforcement, upgrade/migration, regression) over the real bootstrapped repos. No P0/P1 survived independent verification — the calibration demotion, dual ignore-layers, init-skill sparse handling, and git-committed artifacts each downgraded a claimed P1 to P2 — but eight real defects ship fixed, each regression-pinned. Full unit suite 3,622 passing.

Fixed

• Multi-lens turn-end review could exceed the Stop hook's wall-clock cap. When enforcement.multi_lens_review is on, the correctness and duplication lenses ran sequentially, each spawning a reviewer with its own ~45s budget — a worst case of ~90s that the 55s stop-backstop timeout would SIGKILL, losing the review while the per-session budget was already spent. run_lenses now runs the lenses concurrently, so the pass costs the slowest single lens (~45s), staying under the cap; per-lens fail-open and lens-ordered synthesis are preserved.
• Ignore directives using the full rule name did not silence the advisory for naming / file-naming / inheritance violations. The lint-engine gates checked only the short token (naming-convention) while the emitted rule — and the name the violation message tells users to copy — is the long form (naming-convention-violation); only import-preference accepted both. The block was already overridable (the block path matches the long name), but the advisory kept firing. All four gates now accept both forms, matching import-preference.
• Inheritance-convention check flagged classes that ARE an established base. A class whose own name is one of the archetype's known bases (e.g. BaseController told to inherit Api::V1::BaseController — itself), or a Rails application root (ApplicationController/ApplicationRecord/ApplicationJob/ApplicationMailer), is now exempt. On a real Rails repo this removed 5 false positives, lowering a calibration fp_rate that had demoted the rule; a genuinely wrong base still flags.
• A leading UTF-8 BOM skewed the runtime dimension snapshot. The regex lint extractors (_extract_typescript / _extract_ruby) missed a first-line declaration when the file began with U+FEFF, producing spurious dimension-mismatch advisories on BOM-prefixed files (the dumper bootstrap path was already immune). Both extractors now strip a single leading BOM.
• %(...) percent-literals with a nested paren were not blanked. A Ruby %(eval(x)) string literal leaked its eval( into the dangerous-sink scan because the bracket-pair delimiter arms were non-nesting, risking an eval-call false positive. The four bracket-pair arms now accept one level of balanced nesting (x % (a + b) modulo is still correctly left alone).
• Refreshing after idioms.md was deleted silently emptied it. A corrupt idioms.md warned and was preserved, but a deleted one wrote the empty template with no warning — surprising for the one user-authored artifact. Re-deriving over an existing profile with idioms.md absent now warns to restore from git before the empty template is committed.
• bundle exec minitest (and a bare minitest) was not classified as a test command, so the Stop-gate test nudge could fire after a passing minitest run. Added a standalone minitest runner pattern alongside the ruby -Itest form.
• profile.summary.md reported "0 rule(s)" for the TypeScript config. count_config_rules only counted a nested rules sub-key; the tsconfig/typescript block stores its settings (strict, target, ...) at the top level, so it now counts those minus the wrapper keys.

Documentation

• Documented CHAMELEON_JUDGE_MODEL (the model the turn-end reviewers spawn; default sonnet) in CLAUDE.md, and corrected a stale "20 tools" docstring in tools.py (the surface is 37 tools).