v6: Same network as v5, trained with the reviewed temporal windows (1202 kept, 417 dropped). Hand F1 0.825 (v5 0.825), row F1 0.879 (v5 0.873), reviewed-set F1 0.856 (v5 0.849), raw flicker 8.7/min, tracker 1.37/min at 97.6%. Strict fp16 check clean; not checked on a Mac or iPhone before publishing.
v5: Stop line only: one output, line (no lead box); FastViT-MA36 + D-FINE neck, no SLSDNet stage. Line F1 0.825, row F1 0.873 (v4 0.836), false lines when none 13.4% (v4 16.1%); 537/537 ops on ANE. Needs an app build with comma-leads (8b344e4) or later.
v4: D-FINE hybrid-encoder neck replaces the FPN (FastViT-MA36 trunk, same recipe): line F1 0.828 (v3 0.816), lead F1 0.863 (0.852), line flicker 8.5/min (13.5); fp16-safe, 544/544 ops on ANE
v3: fp16-safe FastViT-MA36 (fixes v2's Core ML overflow): trained in fp16 with QK-norm attention and an activation penalty; whole model on the ANE. Line F1 0.816, lead F1 0.852, line FP after passing 3.7% -> 0.5% (seed 0)