Report policy evidence coverage and common-set comparisons

[iskandr]

Missing predictions can change vaccine selection even when scores on the shared assessable candidates are unchanged. This adds public coverage summaries and policy comparisons over retained occurrence decisions, preserving unknown raw evidence separately from score filling and acceptance.

policy_coverage accepts explicit occurrence universes and exact model requests, including failed requests with no rows. summarize_policy_coverage reports denominators, reasons, distinct candidates and raw-row unions without multiplying allele-free projections. compare_policy_evaluations retains both memberships and computes raw score differences only on the shared assessable set. Reports use existing typed CSV/TSV persistence and keep policy definitions, contexts and provenance.

A composed fixture and the released Vaxrank 3.36.0 builtin:openvax-v1 consumer workflow show a short-context overlay changing the chosen window while shared assessable scores stay unchanged. The baseline recipe is unchanged. Model availability is explicitly separate from criterion-to-model lineage.

Validation: lint passes; 82 focused and released-Vaxrank checks pass, including CSV/TSV replay and the entire coverage contract under both pandas string modes, missing/failed requests, model/version/MHC scope, historical unnamed-version spellings, conflicting evidence and repeated discoveries. The full single-worker ./test.sh gate passes: 5,108 tests, zero skips and zero unexpected warnings. Version: 5.93.0.

Closes #457. Establishes the coverage reporting path for the separate shared cleavage contract in #288.

Closes #461: construct the duplicate-version fixture directly from records to avoid pandas all-null concat deprecation; warnings remain errors.

[coveralls]

coverage: 93.238% (+0.003%) from 93.235% — feat/policy-coverage into master