release: v0.3.0 — gather/scatter, async Metal, fused Adam, linalg, einsum, f64, param groups (#23, #25–#30)

[vyncint]

0.3.0 — the "what oxmega needs" milestone, plus the pins triage

Closes #23, #25, #26, #27, #28, #29, #30.

What changed

• pins: upstream drift — termlens@0.8.0 cuda-oxide@97f8b2b7 #23 pins. termlens 0.6.1 → 0.8.0 in its own commit (PTY suites and both 100-iteration stresses unchanged). cuda-oxide stays at a766fc26: upstream HEAD requires nightly-2026-08-28, reconverge 0.4.0 is a rustc driver built on nightly-2026-04-03 and its conformance/PIN still verifies a766fc26. The pins watch now compares against what the pinned reconverge tag records instead of upstream HEAD, so a drift issue means something this project can act on. Run locally against the current pins: no drift.
• Native gather/scatter on Metal and CUDA (index_select/index_add round-trip through the CPU) #25 gather/scatter. index_select/index_add run on Metal and CUDA (gather_dim, scatter_add_dim: one thread per output element scanning the index list — deterministic, adds in index order like the CPU, no atomics). The fallback now moves every operand through the CPU round-trip; moving only the target failed the narrow VJP with a DeviceMismatch on CUDA (oxmega's k-DPP loss).
• Small batched linear algebra: det/log-det, cholesky, eigh, plus diag/trace/eye constructors #26 linear algebra. eye/eye_on, diag, diag_embed, trace, cholesky (Murray-2016 VJP, host f64), logdet, det, eigh (Jacobi). cholesky/eigh are backend primitives with a CPU implementation and the index-op fallback contract; non-PD is a typed error naming the batch.
• f64 tensors (or f64 accumulation) for research metrics and set-likelihood normalisers #27 f64. DType::F64 on the CPU end to end (storage, every primitive, autograd, to_dtype); GPUs stay f32 with typed errors.
• Asynchronous dispatch / op batching on Metal and CUDA: per-op launch cost dominates small workloads #28 dispatch. Asynchronous Metal dispatch (encode into an open command buffer, commit every 64 ops or at a host read, wait on every outstanding buffer — Metal finishes them out of commit order) and a fused adam_step kernel on both GPU backends. register_default() is idempotent: a stateful backend must not be replaced mid-flight. ADR-0008.
• Per-parameter-group weight decay / learning rate in the optimizers #29 optimizer groups. ParamGroup { params, lr, weight_decay }, with_groups on every optimizer, groups_mut() for schedules; single-group constructors produce identical updates (asserted).
• Higher-rank matmul broadcasting (rank 4+) and einsum #30 matmul/einsum. Any rank ≥ 2 with NumPy batch broadcasting, lowered onto the unchanged rank-3 backend contract from recorded view ops; einsum for one- and two-operand contractions.

Measured

• Metal, oxmega device-bench (M4 Pro, 100 epochs × 3 stacked seeds, back-to-back with the published 0.2.0, three rounds): 19.4–20.1 s → 7.4–7.6 s; CPU 6.0–6.4 s throughout. The GPU is within 1.2× of the CPU on a workload built for the CPU; it does not overtake it, and LIMITATIONS says so.
• A10G (sm_86, CUDA 13.2): kernels.ptx regenerated (8 entries); parity 11/11 incl. the new index and fused-Adam tests; full workspace green with CUDA registered; compute-sanitizer memcheck 0 / racecheck 0 (whole suite, single-threaded; one parallel run showed 1 hazard that five reruns did not reproduce) / synccheck 0; train --device cuda --epochs 3 reaches loss 0.6693 (0.2.0's trajectory).

Lessons recorded

• Waiting on the newest committed Metal command buffer is not enough; Metal finishes buffers out of commit order.
• Re-registering a fresh backend from every test silently orphans encoded work once dispatch is asynchronous.
• A downstream consumer (oxmega) found the fallback bug that the parity suite never exercised: the narrow VJP on a device tensor. It is a parity test now.