feat(dataset): ploidy-1 unphased union view (#222)

[d-laub]

Summary

Adds an opt-in unphased_union read-time view that folds a diploid dataset's two stored haplotypes onto a single haploid sequence — the union of called ALTs per (region, sample) — for haploid somatic modeling (issue #222).

• New flag Dataset.with_settings(unphased_union=True) on the Haps reconstructor. Stored genotypes stay diploid on disk; this is purely a read-time view.
• Under the flag: ds.ploidy == 1, n_variants(...).shape[-1] == 1 (naive sum across haplotypes, int32-preserved, no dedup), and "variants" / "variant-windows" decode at ploidy 1.
• The core fold lives in get_variants_flat: a pure offset re-grouping (row_offsets[::ploidy]) with no sort and no data movement — safe because the downstream consumer is permutation-invariant. A content-level test confirms the union row is exactly hap-0's calls then hap-1's, with nothing dropped or duplicated.
• Phased "haplotypes" / "annotated" output is rejected under the flag (guarded in both with_seqs and _check_valid_state, covering both orderings).
• Removes the retired, order-dependent germline-CCF inference path (infer_germline_ccfs_ + _infer_germline_ccfs) — unused ~1 year and the only consumer that assumed start-ordering.
• Documents the option in the genvarloader skill (SKILL.md).

Test Plan

• New suite tests/dataset/test_unphased_union.py (15 tests): flag plumbing + genotypes guard, ploidy==1, n_variants fold + int32 contract, both phased-output guard orderings, variant-windows ploidy-axis collapse, union count == sum over haplotypes, union content == hap-0∥hap-1, ragged-variants path, toggle-off restores diploid, AF-filter composition.
• Regression: test_flat_flanks.py, test_flat_variants_type.py, test_flat_mode_equivalence.py, test_rag_variants.py pass (the eff_ploidy == ploidy off-path is unchanged).
• Full suite: pytest tests/dataset tests/unit — 528 passed, 21 skipped, 2 xfailed.
• ruff check python/ clean; pyrefly typecheck 0 errors.

Acceptance criteria from the spec (all covered by passing tests): n_variants(...).shape[-1] == 1, variant-windows P=1 layout, union count == sum over haplotypes, haplotypes/annotated raise under the flag, ds.ploidy == 1.

Scoping note: this implements only the with_settings entry point (not Dataset.open(unphased_union=...)), per the plan — open promotes to the "haplotypes" default which the guard rejects, making it a footgun. Open-parity can be a follow-up.

🤖 Generated with Claude Code