Route VCF/PGEN writes through genoray _dense2sparse_with_length

[d-laub]

Summary

• Route VCF/PGEN dataset writes through genoray 2.7.0's _dense2sparse_with_length, so they produce per-haplotype-minimal sparse genotypes — identical to what the SVAR write path already produced via read_ranges_with_length. Previously VCF/PGEN applied plain dense2sparse to the over-extended window, so every haplotype kept all extension-tail variants (over-extending haplotypes with few/no deletions).
• Fix extend_to_length=False, which was a silent no-op for VCF/PGEN (_write_from_vcf/_write_from_pgen never forwarded it to their chunk generators, which always called _chunk_ranges_with_length). It now genuinely disables extension via the non-length chunking APIs (VCF.chunk / PGEN.chunk). This makes the code match the existing gvl.write docstring and SKILL.md.
• Compute each region's chromEnd (max_ends) from the furthest retained variant via v_ends = POS - clip(ILEN, max=0), matching _write_from_svar exactly, so all three input paths (VCF, PGEN, SVAR) agree.
• Bump genoray dependency to 2.7.0 (pixi.toml ==2.7.0, pyproject.toml >=2.7.0,<3).

Implementation refactors the two near-duplicate generators to assemble each region's full dense window across genoray memory-chunks and delegate conversion to a shared _window_to_sparse helper (dispatch between _dense2sparse_with_length and dense2sparse). Empty regions and chunk/index alignment are guarded in both no-extend paths.

Design + plan: docs/superpowers/specs/2026-05-30-dense2sparse-with-length-design.md, docs/superpowers/plans/2026-05-30-dense2sparse-with-length.md.

Test Plan

• tests/integration/dataset/test_vcf_pgen_svar_parity.py — acceptance: VCF and PGEN sparse output (start/ilen/alt) is byte-identical to SVAR across all region×sample pairs (with a non-emptiness guard so it can't pass vacuously).
• tests/unit/dataset/genotypes/test_window_to_sparse.py — _window_to_sparse trims per-haplotype under extension and keeps carried variants without it.
• tests/unit/dataset/test_write_no_extend_empty_region.py — extend_to_length=False write no longer crashes on variant-free regions.
• Full suite: 494 passed, 11 skipped, 2 xfailed (only the missing-local-data 1kg slow tier is excluded).
• Edge cases probed in review: multi-memory-chunk windows, insertion/deletion boundaries, reverse strand, jitter-expanded coordinates — all parity-consistent.

Note: the .gvl ground-truth datasets are gitignored and regenerated by pixi run -e dev gen (which the test task depends on); per-haplotype-minimal output means they must be regenerated before running the parity test locally.

🤖 Generated with Claude Code