feat(stark): full GPU trace residency — skip the round-1 D2H#799

Draft

MauroToscano wants to merge 7 commits into

feat/gpu-constraint-evalfrom

feat/gpu-trace-residency

Commits on Jul 9, 2026

feat(math-cuda): device row-gather kernels for R4 opening values
MauroToscano
committed
spike(stark): Phase 4 Stage 2 — device-gathered R4 opening values (cross-checked)
MauroToscano
committed
test(stark,prover): counter-guard the Stage 2 device-opening gather
MauroToscano
committed
spike(stark): Stage 3 step 1 — host_trace_empty flag + hard-abort guards
MauroToscano
committed
spike(stark): Stage 3 step 2 — device-only gate skips the round-1 trace D2H
MauroToscano
committed
bench(stark): continuation mode for bench_abba.sh (two-ref GPU residency A/B)
MauroToscano
committed
fix(stark,math-cuda): review fixes — harden device_only gate + inference + gather fallback
MauroToscano
committed