Chameleon

Turn-end review gets a semantic duplication pass and a deterministic test-integrity advisory; the auto-pass router now grades each change into a complexity tier and records it; and an opt-in multi-lens turn-end review lands behind a default-off flag. The duplication change is the first one measured causally on real repos: a powered A/B (46 mined tasks, two passes, n=44 paired) shows it cuts the duplicate/fail-to-reuse rate from 86.4% to 67.0%, paired bootstrap 95% CI [+0.08, +0.31] (stable across seeds), sign test p=0.0072 — judge-free.

Added

• Semantic duplication at turn-end. The turn-end duplication advisory previously matched only byte-identical (body-hash) re-implementations; it now also runs the name-token + signature-shape semantic engine (the same one /chameleon-pr-review uses), so a helper re-implemented with a different body but the same intent is caught and the existing one named for reuse. Gated for precision: a body-identical match always qualifies, a name-only lead must share at least two domain tokens (DUPLICATION_SEMANTIC_MIN_SHARED_TOKENS, default 2), which cut name-overlap noise ~53% on a real-repo sweep. Within-session body-hash detection is preserved.
• Test-integrity advisory (enforcement.test_integrity_review, default on). At turn end, when a turn changed live source AND weakened tests (added skip markers, dropped assertions, net test deletion — the deterministic signals the auto-pass router already computes), it surfaces a one-line advisory naming what was weakened. Deterministic, zero model spawn; advisory only, never a block; per-session digest-deduped so it does not re-nag.
• Complexity tier on the auto-pass verdict. classify_complexity_tier grades a change easy / medium / hard / complex from diff facts alone (size, new files, cross-file blast, security surface) — structural, independent of whether the change is clean. The tier rides on get_autopass_verdict, is rendered by /chameleon-pr-review, and is recorded in the signed review ledger so per-tier review-clean rates can be tracked over time.
• Multi-lens turn-end review (enforcement.multi_lens_review, default OFF — opt-in). When enabled, one coordinated pass runs the correctness and duplication lenses together (no mutual single-spawn defer) and merges their findings through lens_synthesis (cross-lens agreement surfaces; a lone lens only at high confidence). Off by default because it lifts the per-turn reviewer-spawn budget above one; measure in shadow before enabling. Honest status: not yet shown to add over the base gates — ships dormant for measurement.

Changed

• Effectiveness convention scorer is now delta-based. It lints each changed file's baseline version too and counts only the violations the change INTRODUCED, instead of charging a session for the pre-existing violations in whatever file it touched (which confounded the A/B by file choice).

Tooling (local eval, not shipped to plugin users)

• Statistics module (tests/effectiveness/stats.py: Wilson lower bound, Cohen's kappa, paired cluster-bootstrap CI) so effectiveness claims report confidence bounds, not bare percentages; a dup task tier with a 46-task mined-and-verified duplication corpus; a paired-preference CI in the run report; and read-only PR-outcome / tier-distribution measurement scripts that correlate the router's verdict with real review-comment history.