Phase B: Dimension-sweep benchmark - Apple GPU vs CPU crossover #77

Context

Measurement phase of the epic. MPS lost on 32-channel data (#63: 0.51x) because of
fixed per-op Metal dispatch overhead, which amortizes as tensors grow - 32 channels
was the wrong regime to judge MPS. This phase finds the crossover where Apple GPU actually
wins. Runs last so it can benchmark the Phase A float32 path (and Phase C MLX if landed).

Approach

• Extend benchmarks/benchmark_gpu.py with a dimension sweep: channel count
  (32/64/128/256), block_size, and n_models.
• Compare MPS-float32 (and MLX if Phase C landed) vs CPU float64/float32 - warmed, min of
  repeats, CUDA-synchronized.
• Identify and document the crossover point(s).

Data note (timing, not correctness)

Real sample EEG is 32 channels; higher channel counts are produced by tiling/replicating
real channels purely to size the tensors. This is a throughput benchmark, not a
correctness test - the NO-MOCKS rule governs correctness tests, and no correctness claim
is made from tiled data.

Acceptance

• Crossover curve documented in .context/issue-63/perf_findings.md (or a new perf note).
• benchmark_gpu.py runs the sweep behind a single CLI flag.

Dependencies

Blocked by Phase A (needs stable float32); benefits from Phase C (compare MLX).
Research: .context/mps_pathways.md Pathway B.