Admit patient CTA expression with source provenance and configurable MAGE exclusions

[iskandr]

Patient CTA design currently requires callers to assemble antigens themselves. This starts the expression-first workflow in #584 with an adapter that intersects declared patient gene/transcript expression with direct OncoRef CTA admission and returns auditable VaccineAntigen sources, without VCF/BAM.

The adapter uses Topiary's expression loader, retains original feature IDs and input SHA-256, and records missing values, non-CTA features, excluded targets, unresolved transcripts and held-out admission decisions. Gene measurements select the OncoRef canonical reference transcript without assigning abundance to that isoform; transcript measurements select only the measured isoform and reject version mismatches. Identical proteins remain separate source occurrences. Default target eligibility excludes MAGE* except MAGEA4; this never changes the full CTA negative-self-reference exclusions.

Policies, expression measurements, OncoRef reference resolutions, source sequences and decisions persist through allowlisted native JSON. Saved admission replay needs no upstream queries or expression file. Existing peptide and mRNA builders consume the admitted antigen in a synthetic integration smoke. Requires Topiary >=5.94.3, which fixes short expression tables at the shared loader. Bumps Vaxrank to 3.39.0; the OncoRef pin is unchanged.

Related: #584. This PR does not close the complete workflow issue.

Validation

• ./lint.sh: passed.
• Focused CTA suites: 64 passed, including empty, single-feature and multi-feature gene/transcript CSV and TSV inputs; no expected failures.
• Existing peptide and mRNA builder integration passes using a synthetic predictor and a constant Topiary DSL score; no biological binding outcome is asserted.
• Full ./test.sh: 2,213 passed, 349.23 seconds; no expected failures.
• pip check: passed.
• Real-reference smoke with synthetic expression: OncoRef 1.8.206 / GRCh38 Ensembl 93 admits PRAME (509-aa canonical protein) and round-trips the result. MAGEA4 remains an explicit unresolved-transcript decision because the selected older annotation lacks OncoRef's canonical transcript; no alternate isoform is guessed. This is a reference-resolution smoke, not a patient-data validation.

Remaining workflow in #584

• Wire expression intake into the shared CLI prediction, Topiary filtering/ranking, native dataset, report, peptide and mRNA paths; add end-to-end gene/transcript CLI fixtures and real-data smoke.
• Integrate Hitlist presentation observations with positive MS modality, complete source/assay/occurrence provenance and distinct measured/inferred HLA assignments. Non-MS fluorescence/structural assays enter the MS evidence partition and CTA vaccine support pirl-unc/hitlist#644 tracks the upstream modality problem.

Scientific semantics follow Salmon's transcript-abundance model (https://doi.org/10.1038/nmeth.4197) and Ensembl's representative-transcript definition (https://www.ensembl.org/info/genome/genebuild/canonical.html). The example expression threshold in the design document is a caller choice, not a universal cutoff.

Prerequisite: openvax/topiary#487 fixes openvax/topiary#486. The full prerequisite gate passed 5,469 tests with 20 optional skips. A separate high-depth RNA performance concern found during that gate is tracked at openvax/isovar#463. No read cap or Isovar workaround is part of this PR.