[iskandr]

Add a small offline regression from the original public SiHa-1 RNA data (GSM4785628 / SRR12650977), replacing no existing synthetic tests. The slice retains all 46 SAM records for 23 selected query names, including mates, qualities, flags and tags. Original-record hashes, the complete source header, alignment-reference and ORFdata catalogue pins, extraction recipe and data-usage attribution are included. Version 1.64.1.

The normal isovar viral command reconstructs seven paths from this slice and recovers the already reported 344-nt local RNA / 98-aa E7 coding prediction with its sample-specific difference. CDS coverage is tiled; no single segment spans it, RNA strand and transcript completeness remain unknown, and no protein-expression evidence is inferred.

The unchanged historical 26-path / 71-target-fragment replay is retained separately, including human-competitive placements. The bounded slice deliberately excludes query names with outside-target placements and cannot estimate whole-library specificity, abundance or absence. Its reduced reference dictionary uses explicit custom-reference validation and preserves the original missing-M5 limitation.

Validation: lint, the two focused source-integrity/CLI tests and all 6,623 tests pass. The extraction recipe independently recreated the fixture and expected hashes. No reconstruction algorithm changes.

Closes #531. Provides original-read provenance for the downstream openvax/vaxrank#656 follow-up.

[iskandr]

Reviewed the fixture, extraction recipe, source receipts, library/CLI regression and documented evidence limits. No reconstruction algorithm changes or blocking findings remain.

Lint and the full suite pass: 6,623 tests, six existing collection warnings. The focused integrity/CLI tests pass in 1.54 seconds. Every retained SAM record matches its original hash, both mates and original qualities are retained, and the full original header and historical 26-path/71-fragment replay remain separate from the 23-fragment slice. Custom subset-reference replay preserves the missing-M5 limitation and exact reference/catalogue pins.

The original 344-nt RNA / 98-aa local coding prediction is recovered with tiled CDS support and no whole-CDS segment; strand, full transcript and translation remain unproven. A downstream Vaxrank smoke preserved all 13 derived antigen candidates and the complete source result through save/load; the single complete local E7 candidate was not admitted without a tumor-specificity attestation.

CI and clean-master deployment remain required before release.

[coveralls]

Coverage is 92.375% — test/siha-original-rna into feat/patient-background-pairs. No base build found for feat/patient-background-pairs.

[iskandr]

Released Isovar 1.64.1 from clean master after passing CI on Python 3.11–3.14, lint and the complete suite. Verified the live PyPI wheel against the merged source and release tag.

The bounded offline fixture retains 23 original fragments (46 mate records), their original alignment/quality information, source-header and full-input replay receipts, reference/catalogue hashes and a reproducible extraction recipe. The two regression tests recover the existing sample-derived coding hypothesis without treating tiled RNA as a single spanning molecule or observed translation. All 6,623 local tests pass.

The Vaxrank adapter smoke preserves the complete source through native save/load (13 candidates; one complete local E7 coding candidate; no implicit tumor-specificity admission). Downstream ranking/report regression remains tracked separately in openvax/vaxrank#656.