refactor(concat,svar2): clear the three follow-ups from the analytic run planning review

[d-laub]

Closes #411. Closes #412. Closes #413.

The three follow-ups filed during the final review of #414. Independent of each
other; split into one commit apiece.

#411 — fold the readbound-tracks fixture onto a shared store

tests/dataset/test_svar2_readbound_tracks.py duplicated _SVAR2_SLOT_REF /
_SVAR2_SLOT_VCF byte for byte (verified by diff — only the constant name
differed), and rebuilt a .svar2 store with parameters identical to the one
phased_svar2_gvl built inline. The same store was produced twice per session,
by two modules, from two copies of the same inputs.

Extracted as a session-scoped svar2_slot_store fixture that both consume.

The test module uses the fixture under its own name rather than aliasing it to
svar2_store: it lives under tests/dataset/, where tests/dataset/conftest.py
defines a different 3-sample svar2_store that an alias would silently have
bound to. Confirmed with --fixtures-per-test that the tests resolve to
tests/conftest.py:180.

#412 — promote svar2_store_2s to session scope

It was module-scoped because test_fingerprint_detects_mutated_store appended a
byte to a .bin inside the shared store; under session scope that corruption
leaked into every later consumer (17 failures plus a Rust
cast_slice -> OutputSliceWouldHaveSlop panic). That test now copies into
tmp_path and mutates the copy, so the scope can widen.

Measured with --setup-show across the six consumer modules: 6 fixture builds
under module scope, 1 under session scope. Each build runs genoray's conversion
pipeline in a subprocess.

Verified against the full tree, not a scoped run — this failure mode surfaces
in other modules.

The fixture docstring now states the read-only contract and why, so the next test
that wants to mutate it copies first instead of rediscovering the panic.

#413 — hoist np.unique out of the per-batch loop

copy_runs called np.unique(ds_vec) once per batch. On the samples axis
RunPlan.slot_batches computes ds_vec once, outside its region loop, and yields
the same array object for every region — so np.unique re-sorted an identical
array R times (3,734 sorts of a 1.07e6-element array at the chr22 grid).

slot_batches now yields the distinct source indices as a fourth field. It knows
them for free: regions-axis and ExplicitRunPlan batches are each single-sourced,
and on the samples axis the grouping is loop-invariant. That also enables a
single-source fast path in copy_runs that skips building an all-True mask over
as many as _SLOT_BATCH_SLOTS elements.

Measured on the offsets loop, outputs asserted identical to the previous
implementation:

axis	slots	before	after
samples, 2 interleaved datasets	19.2e6	0.268s	0.185s
regions, block concatenation	9.6e6	0.119s	0.049s

The new field is pinned by the existing provenance oracle test, which now also
asserts it equals np.unique(ds_vec) — copy_runs trusts it to choose which file
to read offsets from, so a wrong value would be a silent wrong merge rather than a
crash. The fast path was mutation-checked: zeroing its length computation fails
three concat tests.

Testing

• pixi run -e dev pytest tests -q — 1265 passed, 61 skipped, 4 xfailed, 0 failed
• ruff check / ruff format clean; pyrefly 0 errors (47 suppressed, 596 warnings — non-vacuous)
• api.md in sync with __all__ (MISSING: none); no public API touched, no Rust touched

Not done here

tests/ still holds 7 copies of the same 40 bp reference string. Only the
readbound-tracks module matched both the reference and the VCF, so it was the
only safe fold; the rest pair that reference with different VCFs. Consolidating the
bare constant across 7 modules is a wider change than these follow-ups and is left
alone deliberately.

🤖 Generated with Claude Code

https://claude.ai/code/session_015VxRqNngU7Eg1wdb1aEgD6