[SizzleUnrlsd]

Closes #143. This fixes a soundness bug found while designing enterprise-validators, before release 0.8.0 (#142, held until this lands).

Problem

The refutation accepted a Validator as a proof for every kind, whatever the kinds it declares. A redirect validator guarding os.system(target) refuted the command injection, and the finding was dropped. No bundled model was affected (re.fullmatch declares all kinds). But user-declared validators would each have silently refuted every other kind.

Change

• What a proof covers. recognise() returns, with each proof, the kinds it covers. Built-in checks (string validators, constant allowlists, equality with a constant) cover every kind, as before; a Validator covers the kinds it declares.
• When a proof counts. A proof counts for a flow only when it covers every kind the flow carries into the sink. Otherwise the guard is an unproven one, and the flow is a hotspot.

Verification

• pytest: 1082 passed. ruff check src tests and mypy are clean.
• pytest -m regression under Python 3.12: 19 passed, no snapshot changed.
• The new test, red before the fix, pins four cases:
  • REDIRECT validator on a COMMAND flow: hotspot;
  • COMMAND validator on a COMMAND flow: refuted;
  • COMMAND validator on a COMMAND|SQL sink: hotspot;
  • COMMAND|SQL validator on a COMMAND|SQL sink: refuted.
• Mutation check. Ignoring the kinds again fails that test.