The inference dropped the complete :shape whenever it rebuilt a struct type
(cap) or joined two (join-t/merge-fields), so a vec3 retrieved from a container
or a fn param typed across call sites lost its layout and every field read fell
to the slow descriptor path. Two fixes:
- cap preserves :shape: capping truncates field VALUES below the depth limit but
never the key SET, so the layout is still complete. It also recurses into
fields, so a shaped value nested in a container (a vec3 inside a hit-info)
keeps its own :shape — which is what lets (:r (:normal hit-info)) bare-index.
- join-t preserves :shape when both sides are the SAME complete shape (the
merged struct has the same keys); different shapes still drop it. This carries
the shape through if-joins and the inter-procedural fixpoint's call-site joins.
Result: the ray tracer goes from 22s (R1, correct-but-descriptor-path) to 4.36s
— 2.7x FASTER than the 11.7s no-shape baseline, and ~3x the JVM (was 8.5x), with
byte-identical output. The compounding of cheaper tuple construction plus
bare-index reads across the whole render far exceeds the per-op estimate.
Gate green flag-off, suite 4718, default-path bench even, transparency intact.
e61a175f91