feat: per-string hashing of Ragged containers (md5/sha256/rapidhash)

[d-laub]

Summary

Adds per-string hashing of Ragged string containers via three algorithms, backed by a single rayon-parallel Rust kernel:

• seqpro.rag.hash(rag, algo, *, seed=None) (canonical free function) and Ragged.hash(algo, *, seed=None) (thin delegator).
• algo: "md5" → (N, 16) uint8, "sha256" → (N, 32) uint8, "rapidhash" → (N,) uint64. seed is rapidhash-only.
• Output is a regular NumPy array when strings aren't grouped above the string level, or a single-level Ragged reusing the input's outer offsets when they are.
• Works for both string representations (opaque 'S' leaf and chars |S1 leaf) at ragged depth R=1 and R=2.

Implementation

• Rust (src/hashing.rs): _ragged_hash PyO3 kernel — generic hash_elems<D: Digest> for md5/sha256 (RustCrypto), portable C++-compatible rapidhash_v1 for rapidhash. All compute runs inside py.detach(...) capturing only Ungil slices, per the project's PyO3 convention. Deps digest/md-5/sha2/rapidhash added; registered in lib.rs.
• Python (rag/_ops.py, _core.py, __init__.py): normalizes any string representation to the kernel's (data, delimiters) contract via to_packed(), then shapes the output. Follows the free-function-canonical / method-delegator idiom (mirrors reverse_complement, concatenate).
• Docs: skills/seqpro/SKILL.md "Hashing strings" subsection + quick-ref row; CLAUDE.md records the free-function/delegator idiom convention.

Testing

tests/test_ragged_hash.py — 22 tests: crypto byte-for-byte vs hashlib, rapidhash identity/seed/determinism, both string representations × R=1/R=2, grouped→Ragged with correct outer offsets, leading-fixed-dims reshape path, gathered/unpacked input, empty container, and all error paths (numeric, record, unknown algo, seed-on-crypto). Full suite: 612 passed, 2 skipped, 2 xfailed, 2 xpassed. cargo fmt/clippy and ruff clean.

Developed task-by-task via subagent-driven development; final whole-branch review (Opus) verified GIL detachment, rapidhash v1 portability/determinism, and the unified delimiter contract across all four input shapes — ready to merge, zero Critical/Important findings.

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will not alter performance

⚡ 2 improved benchmarks
❌ 3 regressed benchmarks
✅ 9 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	test_bench_ragged_cres	1.2 ms	1.8 ms	-32.3%
❌	test_bench_baseline_ragged_short_alleles	2.2 ms	3.2 ms	-30.4%
❌	test_bench_baseline_ragged_cres	6.1 ms	7.3 ms	-16.14%
⚡	test_bench_ragged_flanked_alleles	3 ms	2.2 ms	+35.72%
⚡	test_bench_dense_batch	2.9 ms	2.4 ms	+21.72%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing feat/ragged-string-hashing (4e42def) with main (dbe83ae)}