Preserve fixture provenance and native coordinates; add audited evidence (0.15.0)

[iskandr]

Bundle subsetting now preserves the original headers, acquisition receipts and recorded selection outcomes. Historical retrieval cases now expose their native BAM coordinates and reference dictionaries, including the GRCh37/rCRS mitochondrial case, so callers can reject incompatible merges. New fusion requests use a versioned panel with explicit retained sides.

• Fixes subset_bundle: carve a library's members out of a shared bundle, keeping original headers and receipts #90: subset_bundle plus test-data make --from BUNDLE --member PREFIX, verified against all 345 Isovar members of openvax-v2.
• Fixes The sv-regressions-v1 fusions give orientation but not retained sides #96: sv-regressions-v2 adds the reviewed sides for the three Isovar fusions; published v1 panels and bundles remain unchanged.
• Fixes shared-v1 retrieval cases mix genome builds and contig naming, so merging them silently yields zero reads #99: separately versioned native metadata for all 49 retrieval cases, checksum-verified assets, exact contig names, mitochondrial references, and a merge compatibility guard. All 49 indexed native queries were verified against the original archive.
• Partially addresses Resolve overlapping RNA evidence under two Tempus sample labels before treating them as independent libraries #100: committed regional BAMs and receipts reproduce 40,684 matching core records and 18,971 shared names. File selections can reject known overlapping products or keep one explicit preference while recording exclusions. The published FASTQ catalogue explicitly identifies KCVBE1UI1P as a duplicate delivery of ALMY2X4KMV; corresponding mate objects have matching sizes and multipart ETags. These source receipts and metadata are now pinned too. The existing T0-versus-T1 specimen dispute remains unresolved and is flagged on the RNA count source; specimen documentation is still needed before relabeling it. Issue Resolve overlapping RNA evidence under two Tempus sample labels before treating them as independent libraries #100 remains unresolved pending specimen documentation.

The comprehensive regression corpus includes the union of 51 vaccine variants, all 7,474 variant/BAM support cells, 12 annotated frameshifts, all 637 structural candidates, and four DNA-supported SNVs with zero alternate support in the 15 tested tumor RNA products. Low coverage, BTN3A3's separate blood-RNA observation and NR2F2's historical RNA support remain explicit. Independently recounted unsampled control BAMs are available in the local evidence delivery archive.

The larger balanced openvax-v3 read build is paused for disk space and remains an unpublished candidate. Its completion and disk-use improvements will follow in a separate PR. Completed source selections are checkpointed; the candidate permits 1,000,000 records per supplementary region and 2,000 supplementary regions per source. This PR does not change the published openvax-v1/v2 bundles or certify the unfinished v3 read archive.

Validation: 377 tests passed; Ruff, strict MkDocs and package builds passed. Real-bundle subsetting preserved every selected record, original header and receipt. All 140 independent control recounts passed, including 60 tumor-RNA negative checks. The committed overlap and coordinate regression fixtures run offline.

Review fixes: both control audits bind source URLs and BAM/index hashes to acquisition receipts; compact carry-forward preserves placed-unmapped records; packaged RNA-support statements follow the independent recount results.

Release: osteosarc 0.15.0. Python 3.9 support remains. The full v3 read archive is not part of this release.