Cartridge genotype plane: persist population genotype priors on a cartridge

[MuteJester]

Summary

Adds a donor-population germline prior ("population genotype model") as a top-level cartridge plane (DataConfig.genotype_priors), so Genotype.sample(cfg) can draw per-individual genotypes from a cartridge-authored prior instead of a uniform synthetic one. Follow-on #2 of the genotype roadmap (after V1 #5 and population priors #6).

Python-only — zero engine_rs/ changes. The plane is consumed entirely in Genotype.sample, which already lowers a fully-specified genotype through the existing SampleGenotypePass; population novel alleles reuse the existing effective-refdata injection. No Experiment API change.

What's included

• genotype_priors.py — PopulationGenotypeModel + PopulationNovelAllele: per-segment/gene allele frequencies, per-gene haplotype-deletion probabilities, chromosome weights, population novel alleles, and identity (model_id/source/version/description). Catalogue-free validate() (shape, numeric bounds, DNA, ChainType-normalized D-on-VJ rejection) + canonical content_checksum().
• PopulationGenotypeModel.from_genotypes(...) — pure estimator from observed Genotypes: per-chromosome allele counts (homozygous=2 / hemizygous=1), deletion prob = deleted_haplotypes / (2·n_subjects), pseudocount on catalogue alleles only, with completeness / copy-number / subject-id / input-type guards.
• DataConfig.genotype_priors — fifth top-level plane; legacy-pickle fallback; None-pop checksum shim (unused plane → byte-identical cartridge checksum; a real plane → distinct identity); models.genotype_priors manifest block (robust against malformed direct-attached planes).
• Genotype.sample — per-input source resolution with use_cartridge_priors opt-out; candidate-vs-carried novel injection (include_cartridge_novel_alleles); genotype-level prior_provenance + to_metadata() (base + effective refdata hashes); order-independent draws.
• Builder — set_genotype_priors (catalogue-aware + functional novel validation) and chainable estimate_genotype_priors.
• Docs — new "Population genotype models on a cartridge" guide section; all examples executed live.

Invariants

• No-genotype-prior cartridges keep a byte-identical compute_checksum(); a real plane is part of cartridge identity.
• Plane novels are candidate alleles for the draw, but only carried novels reach effective_dataconfig() / AIRR projection — uncarried candidates never leak.
• Backward compatible: no plane attached ⇒ behavior unchanged.

Test plan

• New tests/test_genotype_cartridge_plane.py (45 tests): validation, checksum invariant + canonicality, serialization round-trip, manifest, sampler provenance/opt-out/mixed-sourcing, candidate-vs-carried novel injection, estimator statistics + guards, builder set/estimate, input-type hardening, end-to-end truth-call carriage.
• Full Python suite green (2425 passed).
• No engine_rs/ changes (compiled engine byte-identical).
• Reviewed across four rounds (spec-compliance + adversarial API/statistics + two external critic passes); 19 findings triaged and fixed with regression tests.