Split from #682's composition arc (ponder-lab#617): recovering the degraded ? of float32 members requires the type-feed criterion to cover known-dtype unknown-shape seeds, not just pure-⊤ ones, so the feed can compose shapes for results whose dtype the generator already proved.
The widening is implemented trivially (drop the dtype guard in the engine's feed loop) but regresses two vendored probes, testCollectionProbeVendoredConv1d (the pinned reshape loses its state entirely) and testGatherElementsAlongRow (a variable drops): suppressing a known-dtype seed couples into engine-side pin computations that read seeded state. The bisect isolating the widening as the sole cause is recorded on ponder-lab#617. Untangling that coupling (or restoring more than empty-state destinations) is the work here.
Split from #682's composition arc (ponder-lab#617): recovering the degraded
? of float32members requires the type-feed criterion to cover known-dtype unknown-shape seeds, not just pure-⊤ ones, so the feed can compose shapes for results whose dtype the generator already proved.The widening is implemented trivially (drop the dtype guard in the engine's feed loop) but regresses two vendored probes,
testCollectionProbeVendoredConv1d(the pinned reshape loses its state entirely) andtestGatherElementsAlongRow(a variable drops): suppressing a known-dtype seed couples into engine-side pin computations that read seeded state. The bisect isolating the widening as the sole cause is recorded on ponder-lab#617. Untangling that coupling (or restoring more than empty-state destinations) is the work here.