[RFC] perf(recursion): verifier optimizations — paired Merkle opening, keccak direct permutation, scratch buffers#706

Draft

Oppen wants to merge 75 commits into

mainfrom

Commits on Jun 23, 2026

Commits on Jun 24, 2026

perf(stark): batch-invert composition denominators + hoist z^N_parts out of query loop
Oppen
committed
perf(crypto): direct keccak state absorb for small field-element leaves
Oppen
committed
perf(stark): batch-invert all 292 trace denominators across all queries
Oppen
committed
perf(stark): remove unused proof_sym Merkle paths from proof format
Oppen
committed
perf(stark): single-pass OOD inner product for height=2 common case
Oppen
committed
perf(crypto): streaming keccak leaf hash without intermediate byte buffer
Oppen
committed
perf(stark+executor): Fp3 FMA precompile — fused multiply-add in one ecall
Oppen
committed
perf(stark+executor): scalar×Fp3 FMA precompile + eliminate evaluations Vec
Oppen
committed
perf(stark): fma for trace_term, h_terms, boundary, and transition sums
Oppen
committed
perf(stark+executor): scalar-Fp3 dot product precompile — n FMAs in one ecall
Oppen
committed
perf(stark+executor): Fp3-Fp3 dot product precompile + ext-column batch
Oppen
committed
perf(stark+executor): b_terms via Fp3Dot + combined coeff precompute cleanup
Oppen
committed
perf(crypto): little-endian leaf hash protocol — eliminate swap_bytes per element
Oppen
committed
perf(crypto): raw LE leaf hash — skip Goldilocks canonical_u64() per element
Oppen
committed
perf(crypto): add write_bytes_le override to Fp3 element (zero-copy)
Oppen
committed
perf(stark): fma in FRI fold — replace Fp3Add with Fp3Fma ecall
Oppen
committed
perf(crypto): raw LE in Fiat-Shamir transcript — skip canonical+swap per element
Oppen
committed