Phase B: cross-platform dimension-sweep benchmark

[neuromechanist]

Summary

Epic #74 Phase B: the benchmark that answers what Phases A and C were building toward -- does an Apple/NVIDIA GPU actually beat the CPU for AMICA, and where? Measures both results (converged LL) and performance (ms/it) for every backend (numpy / torch-cpu / torch-mps / torch-cuda / mlx) on real 70-channel EEG (OpenNeuro ds002718 sub-002), sweeping channel count and n_models with component sharing on/off.

benchmarks/benchmark_dimsweep.py auto-detects the host's backends, records warmed min-of-repeats timing + LL at matched settings, and emits JSON so a Mac run and a CUDA-host run merge via --report. MLX (single-model MVP) auto-excludes from multi-model/sharing configs. Data is fetched from OpenNeuro (public), not committed (README_dimsweep.md; benchmarks/data/ + result JSONs gitignored).

Headline results (single-model, ms/iteration)

ch	mlx	cuda-f32	torch-cpu-f32	torch-mps-f32	numpy
32	21	36	143	162	624
70	25	36	173	255	1350

• MLX is the Apple-GPU win: ~15-25 ms/it, flat across 16-70 ch, ~7x over torch-CPU and faster than an RTX 4090 (CUDA ~36 ms) at EEG scale.
• PyTorch-MPS never wins (162-255 ms/it, at or worse than CPU) -- the Apple-GPU acceleration is MLX, not device="mps".
• Results agree across cpu/mps/cuda/mlx and f32/f64 to ~3 digits on real data (LL ~ -3.216 at 70ch).
• Multi-model has no GPU path yet (MLX MVP is single-model; MPS loses) -- multi-model MLX is the top fast-follow.

Full tables + caveats: .context/issue-77/benchmark_findings.md.

Test plan

This is a benchmark harness (no unit tests / not run in CI). Validated by running it: full single-model sweep on Apple Silicon (cpu/mps/mlx) + RTX 4090 (cuda), and multi-model (m2, m2+share) on Apple Silicon. --report merges the per-platform JSONs into ms/it + LL tables. ruff clean.

Caveats / pending

• MLX vs CUDA is cross-machine (Apple M-series vs an RTX 4090 host), not a same-box comparison -- read as "best Apple-GPU path vs a strong NVIDIA GPU".
• Multi-model CUDA is pending (the hallu SSH session dropped mid-run); single-model CUDA is complete. Re-run --n-models 2 [--share] --backends torch-cuda-* on the CUDA host to fill it in.

Closes #77
Part of epic #74

[neuromechanist]

Review (Sonnet code-reviewer, worktree-isolated)

One valid finding (confidence 84), fixed in 8f26b39; everything else confirmed correct.

Fixed: the numpy adapter didn't pin block_size, so AMICA_NumPy used its default (128) and ran an in-fit do_opt_block block-size auto-tune search inside the timed region -- unmatched vs torch/mlx (which pin 512) and timing pollution that ~doubled numpy's ms/it. Now passes block_size=512, do_opt_block=False. Corrected numpy: 142-622 ms/it (was 310-1350) -> numpy is ~10-25x slower than MLX, not 50-70x. Results (LL) are unchanged (block_size doesn't affect the accumulated sufficient stats), and the headline conclusions (MLX ~7x win; PyTorch-MPS never wins) are unaffected. Findings table updated. (The PR body's numpy column above is pre-fix; see .context/issue-77/benchmark_findings.md for the corrected numbers.)

Confirmed correct by the reviewer (cross-referenced against the actual backends): the numpy LL normalization (m.ll[-1]/(ch*samples) matches the torch/mlx per-sample-per-channel scale), the CUDA synchronize() placement, MLX timing (fit's per-iter mx.eval prevents lazy carryover), the MLX single-model availability gate, the share_iter/share_int handling (sharing fires at iters 5/13 within the 20-iter budget), the --report formatting, and compliance (no emojis/attribution, real data not committed, timing benchmark on channel subsets of real EEG).