Chameleon

Per-edit off-pattern counterexamples. The canonical witness shows the model the
right way to write an archetype; it never shows the wrong way the team has
explicitly flagged. When a team teaches a competing import ("prefer X over Y")
and a real file in the repo still imports Y, that line is now captured and paired
with the witness at edit time as a grounded "do NOT write it this way" directive,
the positive/negative contrast in-context-learning research favors over a
positive example alone. Measured in an isolated agentic A/B (real sessions, no
repo access so the injected block is the only signal): +100pp adoption of the
team's wrapper in both languages (TypeScript and Ruby) when the witness does not
already demonstrate the concern.

Added

• Off-pattern counterexample in the per-edit block (default ON,
  CHAMELEON_COUNTEREXAMPLE=0 disables). A taught competing import whose
  discouraged module is still used somewhere in the repo surfaces that real line
  next to the canonical witness. The signal is conservative: it fires only on a
  TAUGHT competing pair (never auto-derived) with a present usage, so a clean
  archetype injects nothing and the index never fabricates an anti-pattern. The
  artifact (counterexamples.json, trust-hashed) is built at teach time and at
  bootstrap/refresh, never on a hook hot path; the edit-time read is mtime-cached
  and fails open. Rendered outside the imitate-spotlight (a counterexample must
  not be copied) and sanitized + fence-neutralized.
• Every taught off-pattern is kept per archetype. When a team teaches several
  competing imports for one archetype (e.g. winston→logger AND moment→date),
  every still-present off-pattern is shown, not just the last taught.
  counterexamples.json is a list per archetype (schema v2); a legacy v1
  single-row artifact still loads and is normalized on read, so an existing
  profile keeps its counterexample until the next refresh rewrites it.

Fixed

• The counterexample is suppressed when the canonical witness itself imports the
  discouraged module, so the block never contradicts the form it calls "the
  conforming form."
• merge_profiles declines cleanly instead of crashing when an archetypes
  payload maps to lists (the counterexamples shape) rather than dicts; the
  artifact is a regenerable protocol file and is deliberately not routed to the
  merge driver, rebuilt from the merged conventions on refresh.
• The teach-time off-pattern scan no longer misses an import on the largest
  monorepos. The file cap is raised (CHAMELEON_COUNTEREXAMPLE_SCAN_MAX_FILES,
  default 50000) and bounded by a wall-clock budget
  (CHAMELEON_COUNTEREXAMPLE_SCAN_BUDGET_SECONDS, default 10s) instead of a low
  flat cap that could exhaust inside app/ before reaching a discouraged import
  that lives in a peripheral directory. The scan still breaks early on a match,
  so the budget only binds when the taught module is absent.
• An inline chameleon-ignore import-preference-violation that bypasses an
  enforce-mode import deny is now recorded in the override audit. The lint
  suppresses an ignored rule, so the deny gate re-scans the proposed content with
  the directive stripped to recover the bypassed import and record the override;
  previously the bypass was invisible to get_override_audit. The deny decision
  itself is unchanged.