Chameleon

Two fixes. The hooks now pin a Python they can actually run, so enforcement stops
silently switching itself off on machines without a modern interpreter on PATH.
And teaching the profile no longer bounces you to re-confirm your own change.

Fixed

• Hooks pin a dep-capable Python >=3.11 instead of falling through to a stale
  system interpreter. Each hook resolved its own Python via a ladder that ended
  in a blind python3; on macOS that is /usr/bin/python3 (3.9.x, below the
  floor, no deps), so the hook failed and fail-opened — and in enforce mode a
  real violation could pass unblocked with no signal. A shared resolver
  (hooks/_resolve-python.sh) now walks the bundled venv, then python3.13/12/11,
  then uv run (the same resolver the MCP server uses), and only a
  version-probed bare python3 — or surfaces a one-line SessionStart banner when
  nothing >=3.11 resolves. The resolver runs as a subprocess, so a corrupt or
  missing copy degrades the hook instead of aborting it. The merge driver uses the
  same resolver. /chameleon-doctor now reports the exact interpreter the hooks
  pick (command + version + dep status), and a SessionStart banner flags repeated
  hook fail-opens so a degraded session is not mistaken for a healthy one.

Changed

• Teaching no longer stales your own trust. teach_profile already re-granted
  trust after editing idioms.md; the same now holds for teach_competing_import,
  unteach_competing_import, and the deprecated-idiom paths of
  teach_profile_structured (which /chameleon-auto-idiom uses). A teach you ran
  yourself keeps the profile trusted instead of bouncing you to re-/chameleon-trust.
  The guarantee is bounded: trust is preserved only when the profile was already
  trusted (not when it was untrusted or already stale under a teammate's change),
  and the re-grant still runs the injection scan, so a poisoned teach is refused
  and stays stale rather than silently re-trusted.