fix(#176): splice + var_filter parity between open() and with_settings()

[d-laub]

Summary

Closes #176. Three concrete bugs, one refactor:

• Bug 1 — with_settings(var_filter=...) silently failed to propagate to _recon. After .with_seqs("haplotypes"), _recon is a separate Haps instance with the user-chosen kind. __getitem__ uses _recon, not _seqs, so updating only _seqs meant the filter was a no-op for the code path that actually ran. (This is why the reporter's "Path B" appeared to work — it was producing un-filtered haplotypes.)
• Bug 2 — choose_exonic_variants mis-indexed 2-D SVAR offsets: geno_offsets[o_idx] returned a length-n_slices row, not a 2-tuple, producing garbage o_s/o_e and MemoryError / negative-dim ValueError. Mirror the sibling filter_af kernel: geno_offsets[:, o_idx].
• Bug 3 — the regression test from fix: ndim guard on geno_offsets in choose_exonic_variants second loop #170 used the wrong 2-D layout ((total_variants, 2)), which coincidentally made the wrong indexing return 2 elements per row. Rebuilt on the real (2, n_slices) SVAR layout with n_slices > 2 so wrong indexing fails loudly.
• Refactor — Dataset.open now delegates splice_info / var_filter configuration to self.with_settings(...), eliminating the two-code-path divergence that masked Dataset.open(splice_info, var_filter) breaks choose_exonic_variants; with_settings(splice_info, var_filter) works #176 in the first place.

One in-flight scope expansion

After the refactor, Dataset.open(splice_info=...) now reaches _check_valid_state via with_settings, which rejects splicing with the default sequence_type == "variants". Dataset.open auto-promotes to with_seqs("haplotypes") in that case (narrow guard: only when splice_info is set, _seqs is Haps, and sequence_type is the default "variants"). This eliminates a pre-existing silent footgun where the OLD code returned a Dataset in an invalid state until the user called .with_seqs("haplotypes") themselves.

Files changed

File	Change
python/genvarloader/_dataset/_genotypes.py	2-D indexing fix in choose_exonic_variants
python/genvarloader/_dataset/_impl.py	with_settings propagates var_filter to _recon; Dataset.open delegates splice/filter to with_settings; auto-promote to haplotypes when splice_info is supplied
tests/dataset/genotypes/test_choose_exonic_variants.py	Rebuilt 2-D regression on real (2, n_slices) SVAR layout
tests/dataset/test_with_settings_var_filter.py	New unit tests for var_filter → _recon propagation
tests/dataset/test_open_vs_settings_parity.py	New end-to-end parity tests

Test plan

• pixi run -e dev pytest tests/dataset/genotypes/test_choose_exonic_variants.py — both pass on the real (2, n_slices) layout
• pixi run -e dev pytest tests/dataset/test_with_settings_var_filter.py — direct _recon.filter propagation probes pass
• pixi run -e dev pytest tests/dataset/test_open_vs_settings_parity.py — both paths produce identical state and byte-for-byte identical [0,:] output on SVAR-backed data
• pixi run -e dev test — full pytest + cargo, 459 passed, 5 skipped, 2 xfailed
• pixi run -e dev ruff check python/ tests/ — clean

Out of scope

• The min_af/max_af/var_fields recon-rebuild block at _impl.py:427-434 has a separate latent bug (clobbers _recon.kind). This fix deliberately does not touch it.

🤖 Generated with Claude Code