v1.9.98

Bayesian explanation consistency (CAL7, CAL14)

Two explainability fixes so the reported evidence matches what actually drove the
posterior.

• CAL7: grouped evidence was reported as independent. Correlated co-firing
  bindings that share a group (the M365 / GWS indicators, the DMARC policy pair)
  are redundant readings of one fact, so the posterior already reduces them to
  one effective binding. The influence ranking (evidence_ranked), effective
  sample size (n_eff), and evidence_count now use that same contributing set
  instead of the raw fired list. Previously --explain-dag showed each grouped
  binding as a separate influence with its own percentage, and n_eff
  over-counted them, giving too tight an interval. The Evidence line still lists
  every observed binding; only the influence and interval math use the
  contributing set.
• CAL14: declarative absence was hidden from explanations. For a declarative
  node (email_security_policy_enforcing), the absence of an expected public
  declaration moves the posterior, but the DAG renderer printed "no direct
  evidence (posterior follows network priors)" when no binding fired, which is
  false. NodePosterior now carries an absence_informative flag, and the
  renderer explains the posterior as driven by informative absence for those
  nodes.

Gate: full pytest, ruff, pyright (0 errors). The --explain-dag snapshot and
the multi-influence tests were updated to the corrected (group-reduced) output;
multi-influence ranking is now exercised with ungrouped evidence and the top-3
truncation with a direct renderer test.