Add an offline uORF evidence catalogue and mutation-position API

[iskandr]

Variants in 5′ UTRs need a reference for upstream reading frames with actual translation evidence. This adds a versioned, offline human uORF catalogue with exact reference proteins, spliced GRCh38 coding blocks and source-pinned ribosome/MS observations, plus a public Python API and discoverable CLI.

The first dataset has 3,771 upstream models and MS observations for 1,219. Assays, protein intervals, source-specific mapping ambiguity and spectrum reviews remain separate. A genomic-position query reports the uORF residue and overlapping reference peptides; it never promotes native observations into evidence for mutant translation or tumor-specific presentation. TPST1 is a regression example: its observed native peptide covers residues 31–39 and is absent from the 30-aa fusion candidate.

load_uorf_catalog, UORFCatalog.query and UORFRecord.map_genomic_position work offline. isovar uorf-evidence / isovar-uorf-evidence export JSON, TSV and FASTA. The package includes source/license notices, input/output hashes, a reproducible optional source builder and mutation interpretation guidance. Version: 1.46.0.

The complete 2026 curated shotgun tables remain outside this dataset's stated scope; the public export coverage mismatch and concrete follow-up are tracked in #466. Unknown or absent evidence is not treated as a negative result.

Validation: ./lint.sh; ./test.sh (5,826 passed); focused API/CLI regressions; repository-local pip check; installed wheel queried outside the checkout with network access blocked, including both CLI entry points and packaged source/license files.

[coveralls]

coverage: 96.038% (-0.05%) from 96.092% — feature/uorf-evidence-catalog into master