v2.1.14

Assurance: PV2 maintainer-validation loop + inference drift gate

The "Adaptive" assurance-track item: keep the Bayesian CPT numbers honest as the
world drifts. The committable, CI-gated core ships now; the corpus-dependent tier
stays maintainer-local.

• Inference drift gate. validation/drift_check.py fingerprints the Bayesian
  network's CPT-implied marginals (each node's no-evidence prior, all-bindings-
  present posterior, and interval width) deterministically from the network YAML,
  with no corpus. validation/inference_baseline.json is the committed baseline
  (node names and numbers only, no company data), and tests/test_drift_check.py
  gates it: an edit that shifts an implied distribution beyond 0.01 fails until
  the baseline is regenerated with python -m validation.drift_check --update and
  committed, so the shift is reviewed in the same diff. This mechanically enforces
  the CPT-change discipline.
• docs/maintainer-validation.md documents the full tiered loop: the
  committed drift gate (tier 0), the no-data synthetic harnesses (tier 1), the
  public case-study spot-check (tier 2), and the maintainer-local corpus
  re-grounding + firing-rate drift (tier 3), plus how an agent runs it on a
  /schedule routine. Only deterministic / aggregate output is committed; the
  corpus stays gitignored.
• pyright extraPaths now resolves repo-root packages so the gate test can import
  validation.drift_check.

Gate: full pytest (2910 passed), ruff, pyright (0 errors), validate_fingerprint (841), branch coverage 85%.