Osteosarc 0.2.3

The three read consumers currently own separate fixture selection, packing and verification implementations. Move their reviewed historical mechanics into Osteosarc, preserve their declared record membership and scientific expectations, and expose the same versioned panel recipe through each consumer's builder. Varcode adopts the current optional snapshot adapter without adding BAM acquisition to allele-only tests.

This is the fourth PR in the stack: #22 → #23 → #24 → this PR. Addresses #15 and its additional-SV handoff. Version: 0.2.3. Consumer adoption PRs are drafts and must wait for this release: Isovar #351, Topiary #380, Vaxrank #503, and Varcode #489.

• Share historical exact-SAM, regional-corpus and exact-cohort adapters, including pinning, nested manifest updates, safe transport, atomic publication and offline integrity checks. Their documented historical fidelity stays distinct from the lossless BAM bundle API.
• Freeze the five additional SV candidates and three reviewed RNA events in the named panel. Ship the five-candidate recipe plus four original BAM archives, indexes, receipts and saved hypotheses in the sdist: 251 source records, 24 witness/context members, under 4 MiB when exported.
• Retain source-scoped original query intervals, orientation labels and missing-quality uncertainty; no scientific expected outputs are regenerated here.
• Verify acquisition/source identity when using cached subsets, reduce duplicate header metadata, and support Python 3.9 for Isovar.
• Document the complete consumer inventory, regeneration/release sequence, and a practical per-sample BAM panel example.

Validation: 245 core tests on Python 3.12; 228 passed and 17 optional-consumer tests skipped in an isolated Python 3.9 environment. Ruff and strict documentation builds pass. The full 245-test suite passes against the built wheel outside the checkout with network disabled. All three real consumer CLIs produce identical complete record multisets, reasons and source/header provenance from one pinned recipe offline.

Re-review fixes preserve compatibility with full-header bundles produced by 0.2.2, enforce archive size limits before reading metadata, reject mismatched fusion annotations, and publish cohort archives without overwriting concurrently created output. Regression tests and a real 0.2.2 bundle verify those behaviors.

The consumer adoption PRs will be rebased, checked and released in dependency order after Osteosarc 0.2.3 is published. Their previously completed full suites and scientific artifact checks are recorded in those PRs; they will be rerun against the published dependency.