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: 200,000 rows | label (1 float) vs embed (768 floats)

Sequential full pass:
  label only     0.062 s
  full row       0.301 s   (  4.9x more expensive)

Random single-row lookup (median over 10,000):
  label only       23.2 us
  full row         55.5 us
