fp32 is both slower and larger than fp16 on device, and the ANE will not accept fp32 at all, so an fp32 Core ML build is structurally excluded from the accelerator that makes Core ML worth using.
Equivalence was gated with a task metric rather than raw tensor distance: pixel argmax agreement 99.87% across 8 real images. Raw relative L2 on logits reads ~1% for the segmentation models, but that is logit noise which does not change the predicted class per pixel.
Models where fp16 and fp32 genuinely diverge (ssdlite320 86% box agreement, fast-sam 93%) keep both variants.
v0.9.0 and below are untouched.
