chameleon 2.33.3

Fixed

• The corrections-exhausted breaker no longer arms the Stop backstop on a
  non-code file's credential. Once a file is corrected
  MAX_CORRECTIONS_PER_FILE times, advisory feedback is suppressed but a
  deterministic-hard secret still arms the Stop backstop so a credential cannot
  slip in unblocked. Every other arming site and both Stop re-lint branches run
  that secret through block_eligible_on_file(..., language=detect_language()),
  so a credential-shaped token in markdown / config prose (no recognized
  language) stays advisory and never arms the backstop — such a file has no
  inline chameleon-ignore escape and the re-lint drops it anyway. This breaker
  site skipped that gate, so a non-code file that resolved to an archetype (via a
  legacy extension-blind paths_pattern) at the corrections cap could be armed
  inconsistently with the re-lint that then clears it. It now arms only on a
  recognized code language, matching the sibling sites. No behavior change for
  code files.