These are assets the trainer actually loads, not the split file's request: the set produced by tools/init/resolve_consumed_instances.py, whose size matches the run's own Loaded split files line.
Halves are interleaved, so every page mixes TexVerse and synthetic and a systematic difference between them shows up side by side rather than pages apart.
The GT_bake pane is the voxel round trip: the GT mesh re-derived from the very voxel field the trainer reads. The TexVerse half is drawn from assets that already have one; the synthetic half is being baked now, so until that lands its GT_bake panes read not baked yet and a rebuild of this viewer picks them up with no other change.
Compare Full Color against Full Color (no tex normal) to see what the texture normal contributes, and Tex Normal (WS) for the signal the model is trained to predict. Do not read Normal Map (TS) as GT-vs-bake evidence — tangent space is per-unwrap.
Missing on disk: gt 0, gt_bake 29 (the synthetic half accounts for 29 of the gt_bake misses by construction).