[RFC 0005] Testing Serialization Overhaul

[andrewherren]

• Golden fixtures: one serialized model per schema_version and per legacy bracket; load each, assert in-memory state. Old fixtures must load indefinitely and are never modified.
• Per-migration unit tests: vN fixture -> vN+1 fixture for each rung.
• v0 -> v1 heavy coverage: ~one golden fixture per legacy bracket × per platform; exercise the field-presence sub-classifier and the platform-fingerprint fallback.
• Cross-platform fixtures: R-written and Python-written all-numeric model at same schema_version load in both languages to identical state.
• Cross-platform refusal fixture: native-categorical model → cross-platform load errors with offending columns.
• Round-trip identity tests (write → read) for BART and BCF, R and Python.
• Continued-sampling equivalence test: continuing resumes the RNG stream → statistically equivalent to one longer run.
• v > current hard-error test; unsupported-floor / "unknown" rejection test.