build(deps): support genoray 4.x

[d-laub]

Widens the genoray constraint from >=3.4.0,<4 to >=3.4.0,<5 and moves the lock to genoray 4.0.0. 3.x compatibility is retained — the floor is unchanged, so both majors resolve.

Why it's safe

genoray 4.0.0's two breaking changes are on APIs GVL never calls:

• SparseVar2.from_vcf(tune=) removed — GVL never passed tune=.
• SparseVar2.from_vcf_list(max_mem=) became a whole-process byte budget rather than a per-chunk cap — GVL does not use from_vcf_list at all.

GVL leans on a lot of genoray private API, so those were checked specifically. Every private symbol GVL imports is unchanged between the 3.4.0 and 4.0.0 tags:

_types.{POS_TYPE,V_IDX_TYPE,DOSAGE_TYPE} · _contigs.ContigNormalizer · _utils.{format_memory,parse_memory} · _svar2_batch.{MAX_END_SHIFT,_find_ranges_chunked} · _svar2_fields.{StoredField,FormatField,InfoField,_META_DTYPE} · _svar.dense2sparse · _svar._convert._dense2sparse_with_length · exprs.ILEN

_svar2_batch.py is byte-identical across the two tags.

genoray 4.0.0 requires seqpro>=0.21.1,<0.23, which the pinned seqpro==0.22.0 satisfies.

Verification (against genoray 4.0.0)

Check	Result
pytest tests -q (full tree)	1130 passed, 58 skipped, 4 xfailed
pytest tests -q -m slow	6 passed, 3 skipped
cargo-test	134 + 4 passed
ruff check / ruff format --check	clean
typecheck (pyrefly)	0 errors

The first run errored on 3 test_ds_haps_1kg tests, but that was a stale local fixture (untracked, written by gvl 0.35.0 at dataset format v1 vs the current v2) — pre-existing and unrelated to genoray. Regenerated via gen-1kg so the 1kg parity tests actually ran.

Commits

1. build(deps): support genoray 4.x — the constraint widening in pyproject.toml, pixi.toml, docs/requirements.txt (whose genoray>=2.3.1 floor was long stale), plus the lock move.
2. build(deps): align the Rust genoray pin with genoray 4.0.0 — optional, see below.

The second commit is separable

The Rust svar2-codec / genoray_core git pin sat at rev 66ba734 (genoray 3.0.0) — it already lagged the Python package before this PR. Commit 2 moves both crates to rev 1bac5f9 (tag 4.0.0) so the Rust reader and Python writer are built from the same genoray revision, closing a skew that could let an svar2 on-disk format change slip past the Python-writer/Rust-reader tests. No GVL source change was needed.

Cost: genoray 4.0.0 declares tracing and tracing-subscriber as unconditional dependencies rather than feature-gating them, so default-features = false doesn't avoid them. This adds 11 transitive crates GVL never uses (tracing*, matchers, sharded-slab, thread_local, nu-ansi-term, lazy_static, valuable). Gating them upstream in genoray would make this pin free.

Drop commit 2 if that build cost isn't worth it — commit 1 passes the full suite on its own.

Notes

• pixi install rewrote some cosmetic lock formatting (dropped virtual-packages blocks, ./ → .) from a newer pixi version. It's in commit 1's diff and unrelated to genoray; no package other than genoray changed in the lock.
• Left alone: pyproject.toml:11 says >= 3.14 blocked by pyarrow/genoray, but genoray 4.0.0 now allows <3.15 and pyarrow is uncapped >=3.10. The <3.14 cap is likely still correct (conda-pinned numba==0.59.1), but confirming that needs a py314 env — separate change.
• No public API change, so no skills/genvarloader/SKILL.md or api.md update is required; no prose doc asserts a genoray version.

🤖 Generated with Claude Code

[bschilder]

Rejoice! 👼 🎉