Phase 3 close-out: main merge, missing-kernel ports, fused annot/splice haps, seqpro 0.20

[d-laub]

Phase 3 close-out

Brings phase-3-reconstruction to an honest, fully-rust-default state. Closes out the Phase 3 reconstruction + track-realignment migration.

What changed (7 tasks, each parity-gated)

1. Merged origin/main — pulls in the Tracks + max_jitter>0: silent wrong track output (numba) / panic (rust) from interval/query coordinate mismatch #242 intervals_to_tracks sub-query-start clip fix (PR fix(intervals): clip sub-query interval starts in both kernels (#242) #244, both backends) and the fix: SpliceIndexer double-applies sample-subset map (spliced+subset sample scramble) #243 SpliceIndexer subset double-apply fix (PR fix: SpliceIndexer double-applies sample-subset map (spliced+subset sample scramble) #243). The fused tracks kernel inherits the clip fix via the shared intervals::intervals_to_tracks core.
2. Lifted the now-obsolete Tracks + max_jitter>0: silent wrong track output (numba) / panic (rust) from interval/query coordinate mismatch #242 xfails — removed 5 _REASON_242 constants + 10 @pytest.mark.xfail decorators; the max_jitter>0 interval domain is now real, asserted coverage on both backends. The genuine trailing-under-write assume(False) parity guards (a separate, still-live numba-undefined domain) were correctly retained.
3. Rerouted Reference.fetch through the dispatched rust get_reference; deleted the 3 zero-caller _fetch_* numba functions. New tests/parity/test_reference_fetch_parity.py backstop.
4. Fused the annotated-haps path — new reconstruct_annotated_haplotypes_fused (rust): the plain fused kernel + two i32 annotation buffers, reusing the shared reconstruct core. Byte-identical to the composed numba oracle (haps + var_idxs + ref_coords).
5. Fused the spliced-haps path — new reconstruct_haplotypes_spliced_fused (rust): the plain fused kernel minus the diff/offset computation (takes the precomputed splice out_offsets). New tests/parity/test_spliced_haplotypes_parity.py. (The annotated+spliced intersection remains on the unfused dispatched rust core — parity-gated and rust-by-default — with fusion deferred to Phase 5.)
6. Bumped seqpro 0.18 → 0.20.0 and adopted to_numpy(validate=False) at the one guaranteed-uniform read-path site (_reference.py fixed-length branch). The seqpro-core 0.1.0 Ragged layout remains compatible (cargo + parity green).
7. Roadmap honesty pass + full-tree verification — checked off the reconciled Phase 3 items, resolved the ✅-header/unchecked-box contradiction, added a dated decisions-log entry.

Verification

• Full tree, both backends: GVL_BACKEND=rust 932 passed / 12 skipped / 5 xfailed / 0 failed; GVL_BACKEND=numba identical. The 5 xfails are all pre-existing non-Tracks + max_jitter>0: silent wrong track output (numba) / panic (rust) from interval/query coordinate mismatch #242 markers.
• cargo: 88/88. ruff + pyrefly: clean. abi3 wheel: builds.
• Byte-identical parity is the landing gate for every new/rerouted kernel — met across both backends.
• No public API change (__init__.py __all__ unchanged vs origin/main) → skills/genvarloader/SKILL.md needs no update.

Phase 5 carry-forward (non-blocking)

• Thread an explicit output_length through ReconstructionRequest (retire the _out_per == hap_lengths ragged/fixed heuristic).
• Route the three os.environ["GVL_BACKEND"] reads in _haps.py through a single dispatch-aware helper.
• Complete-or-drop genvarloader.pyi (currently stubs only the 3 bigwig functions).
• Fuse the annotated+spliced intersection.

🤖 Generated with Claude Code