(perf): GPU parity — depth-tagged touched-others stack, PoolCheckpointState, lookup memo (CUDA/Metal) - #52

Merged

mgyoo86 merged 10 commits into

refactor/gpu-parity

Jul 10, 2026

Commits on Jul 10, 2026

(refactor): generalize touched-others stack helpers over AbstractArrayPool
mgyoo86
committed
cff21b7
(refactor): adopt PoolCheckpointState in MetalTypedPool; add stack+memo fields
mgyoo86
committed
08489aa
(perf): Metal — depth-tagged touched-others stack + fallback lookup memo
mgyoo86
committed
b02decd
(test): exercise typed Float16 scope in touched-others regression guard
mgyoo86
committed
7e8184e
(perf): CUDA — mirror depth-tagged stack, PoolCheckpointState, lookup memo
mgyoo86
committed
ad2bdd8
(test): cover S>=1 lazy-drain lockstep on GPU pools; retarget stale parity guard
mgyoo86
committed
32035a9
(test): mixed-tuple and R=1 unwind coverage for touched-others; comment hygiene
mgyoo86
committed
28e1863
(fix): GPU poison-fill is best-effort — never throw during rewind
mgyoo86
committed
a85cfc6
(test): auto-manage end-to-end asserts are S-adaptive; scopes return nothing
mgyoo86
committed
b53867d
(test): GPU allocation tests never return pool-backed arrays from scopes
mgyoo86
committed
760d0fd