Retain self-match evidence for vaccine window selection

[iskandr]

Self matching previously exposed one nearest sequence and its first origin, which is insufficient for choosing vaccine windows. Add match_self_peptides / SelfProteome.match_candidates and self_matches_in_windows to retain every match within an explicit same-length Hamming radius, or every exact window occurrence, with all reference origins and caller-resolved exclusions.

Keep supplied presentation observations, confirmed/inferred allele attribution, model records and same-allele coverage distinct. Search completeness and missing-result reasons survive typed CSV/TSV with evidence hashes. Shared CTA/non-CTA sequences retain both origins; oncoref continues to own membership. No match means only no match within the stated search coverage, and presentation does not establish TCR recognition. No predictor, TCR score or new selection recipe is introduced.

The released Vaxrank 3.36.0 consumer fixture loads its actual builtin:openvax-v1 bundle. Self evidence changes the selected window while preserving both intended epitopes, both alleles and all target score. A composed Topiary workflow joins evidence to measurements and verifies explicit unknown handling and exact audit replay. Update the consumer pin and stale baseline/self-evidence documentation; bump Topiary to 5.92.0.

Validation: lint passes; 44 focused and actual Vaxrank consumer tests pass without warnings. The full one-worker ./test.sh gate passes: 5,048 tests, zero skips and no unexpected warnings. PR and master CI are green. Released to PyPI as 5.92.0; wheel/source hashes verified and tag v5.92.0 points to the merged commit. The clean-master release gate also passed all 5,048 tests.

Fixes #455. Fixes #458. Fixes #459. Refs #412 and #224. Follow-up evidence coverage: #457; shared cleavage contract: #288. Broader binding-ranked axes, indel enumeration and automatic evidence acquisition remain in those issues. Primary scientific source: https://www.nature.com/articles/s41541-023-00713-y .

[coveralls]

coverage: 93.235% (-0.03%) from 93.261% — feat/self-match-evidence into master