Chameleon

The turn-end idiom self-review was noisy: every session it re-dumped the full
idioms.md (all archetypes, only reordered) plus the full principles.md, most
of it irrelevant to what was actually edited. This release makes that block terse
without losing coverage, default-ON behind CHAMELEON_STOP_IDIOM_TERSE (set =0
for the old full dump). Verified end-to-end by driving the real hooks against
bootstrapped Ruby (Rails), TypeScript (React), and Python profiles, plus the
Django, Flask, FastAPI, NestJS, and Next.js framework repos.

Changed

• Turn-end idiom review is now scoped, summarized, and de-duplicated against
  what the model already saw. The Stop block (_idiom_review_gate) now: filters
  idioms to the edited files' archetypes plus general/untagged ones, dropping
  other-archetype idioms (B); renders an idiom the model already saw this session
  as a one-line name: summary (C); shows full text only for in-scope idioms not
  yet surfaced this session, so a never-seen idiom is never reduced to a name (E);
  and replaces the full principles.md dump with a one-line pointer, since
  principles are already injected at SessionStart (A). Boilerplate tightened (D).
  New _render_stop_idioms in tools.py.

Fixed

• idioms_shown_names is the honest, per-idiom "the model saw this" signal.
  A new enforcement-state field recording the exact ### idiom names a Tier-2
  block rendered — computed from the char-capped, witness-deduped text the block
  actually showed (_shape_idioms_for_block), not merely "the archetype was seen."
  So an idiom truncated out of the capped block, or one from the deny path (which
  seeds archetypes_seen without emitting idioms), correctly renders full at turn
  end rather than being reduced to a bare name.
• The turn-end idiom parser is now fenced-code aware. A ### line inside an
  idiom's Example: / Counterexample: snippet (a Ruby/shell ### comment, a
  markdown heading) no longer mis-splits into a spurious "idiom" whose gist would
  render the counterexample code as canonical. _reorder_idioms_by_archetypes
  (used on the per-edit hot path and the legacy dump) now shares the same
  fence-aware parser, so the fix covers every idioms.md split site.
• A char-cap truncation never seeds a header-only idiom as "shown." When the
  1500-char Tier-2 cap slices an idiom so only its header (not its description)
  rendered, _idiom_block_names no longer records that tail idiom, so the turn-end
  review shows it in full instead of a gist the model never read. This also drops
  any partial ### header the cut left, closing a rare truncated-name mismatch.
• _merge_states dropped the new field. It rebuilds EnforcementState from
  explicit fields, so the added set was wiped on every save_state, defeating the
  terse path (every idiom rendered full). Now unioned like the other archetype
  sets. Caught by the real-hook driver, not a mock.
• The scope-drift advisory is turn-scoped. It compared the session's
  cumulative changed files against identifiers aggregated from EVERY prompt of
  the session, so one stale early-prompt overlap re-flagged the same files at
  every later Stop — a bare "commit this" turn was scored against the first
  prompt's file names (observed three times in one real session). Intent capture
  now appends an entry for every prompt (empty token lists included, marking a
  request that named nothing), and the advisory reads only the LATEST entry via
  latest_request_identifiers — the turn's governing request. A token-less or
  secret-suppressed newest prompt silences the advisory rather than falling back
  to an older prompt's tokens.
• A principles-only turn no longer burns the once-per-session marker. In terse
  mode the review fires on team idioms, not on the always-present principles, so a
  turn touching no idiom-governed file returns without spending the session's
  review budget — a later governed edit still gets its turn-end idiom review.