feat: native Inspect AI hook for PRML pre-registration (v0.2.0)

[sk8ordie84]

Native Inspect AI hook for PRML pre-registration

Turns falsify-inspect into a proper Inspect extension. Instead of calling the
manual API, you point FALSIFY_PRML at a committed .prml.yaml and run
inspect eval as usual; the hook does the rest.

What it does

• Registered via the inspect_ai setuptools entry point, so Inspect discovers
  it automatically once installed (same mechanism as Inspect's MLflow / W&B
  example hooks).
• At each task end: reads the realised metric from the eval log, checks it
  against the committed threshold, confirms the run's identity (model, dataset,
  task) matches the pre-registration, and writes a *.prml-receipt.json with a
  PASS / FAIL / TAMPERED verdict.
• TAMPERED = the run did not match the pre-registration (e.g. a swapped
  model), so the claim cannot be quietly changed after seeing results.
• Observe-only by default; FALSIFY_PRML_STRICT=1 fails the run on a non-PASS
  verdict (a CI gate).

Scope and safety

• The file-based verify_eval_log API and the falsify-inspect CLI are
  unchanged. inspect_ai stays an optional dependency; the base package still
  imports without it.
• New, tested core helpers: verify_live, verify_observation,
  load_committed_manifest.
• CI now installs the inspect extra so the hook tests actually run.

Verification

• 30 passing unit tests (PASS / FAIL / TAMPERED, strict mode, receipt).
• End-to-end against a real mockllm/model eval: entry-point discovery, a PASS
  run, and a TAMPERED run all confirmed.

Version bumped to 0.2.0. Not yet published to PyPI.

🤖 Generated with Claude Code