Commit: e88caea (dev, post strict-write-dtype) | 2026-07-30 | HDD via WSL2 (ext4, warm page cache)
Machine: x86_64 Linux, 12 cores, 15920 MiB RAM | python 3.10.12, torch 2.13.0+cu130
Data: 50,000 uint8 samples (3, 64, 64), batch=256, epochs=5, torch threads=6

per-epoch raw pipeline      : ['2.55', '2.55', '2.36', '2.49', '2.81'] s (median 2.55)
per-epoch compiled pipeline : ['0.14', '0.14', '0.14', '0.15', '0.14'] s (median 0.14)
one-time compilation cost   : 2.91 s

total over 5 epochs (median):
  raw         12.76 s
  compiled     3.62 s   (3.52x faster)
  break-even: compiled pipeline pays off after ~1.2 epochs
