v2.1.7

Assurance: differential verification of the inference core

The first of the post-2.0 trust-hardening items (roadmap "Trusted" track). An
independent reference cross-checks the Bayesian engine; no engine or output
change, verification only.

• validation/differential_verification.py carries a second inference
  implementation that enumerates the full 512-state joint directly and reads
  each marginal off the normalized sum, with no factor algebra. It re-derives
  the factor construction (correlation-group reduction, declarative absence
  conditioning, soft evidence) from the documented spec, independently of the
  variable-elimination path it checks.
• It sweeps the enumerable evidence space (every node at none/one/all of its
  bindings, ~2.9k configs) plus an exhaustive per-node subset sweep over the
  grouped and declarative nodes. Variable elimination matched naive enumeration
  on every node of every configuration (worst gap 4.95e-05, the engine's
  4-decimal posterior rounding).
• tests/test_bayesian_differential.py anchors the reference to hand-computed
  no-evidence marginals (so the oracle is known-correct, not merely consistent
  with the engine), then holds variable elimination to it on the prior, the
  exhaustive tricky-node sweep, and a strided sample of the full sweep.

Gate: full pytest (2832 passed), ruff, pyright (0 errors), validate_fingerprint (841).