5.7.10

5.7.9

fix: tolerate process shims without env

fix: keep count oracle out of self-host builds

fix: avoid redundant inline local resets

5.4.1: fold reads hex-float constants via f64.parse — Number('0x1p-10…

…22') is NaN and poisoned constant products (jz subnormal pow)

5.4.0

walkN/walkPostN: arrays-only walkers; 89 leaf-blind optimize callback…

…s migrated

Leaves are half of all nodes and most optimizer callbacks guard them out as
their first act — but a self-host build pays a full boxed closure call per
leaf visit (profiled ~10% of self-host compile inside the walkers). walk/
walkPost keep leaf visits for the flat-form token scanners (bare string
opcodes are real instructions: jump/trap detection). Byte-identical output
across the 228-target jz corpus (only self-embedded kernels differ, by the
new source); jz self-host row ~9% faster (29.3 vs 32.6 ms, 3 interleaved
quiet rounds, checksums equal). 5.3.9

brif: br_if pair merge must not speculate a load past its guard — isP…

…ure admits loads and hasTrap misses them (readsMemory now gates the merge; the first br_if of a scan is the second's bounds check); 5.3.8

5.3.7