[LukaszRozmej]

Changes

The eleven zkvm_u256_* shims in src/openvm_intrinsics/openvm_intrinsics.S are only linked in CI (tests/integration/guest.cs) and never run. A wrong funct7 or register order would link fine and then compute the wrong thing. Nethermind's OpenVM guest is about to call zkvm_u256_mul for EVM MUL (NethermindEth/nethermind branch perf/openvm-bigint-arith), so this PR makes CI execute the shims.

• tests/integration/u256.cs (new) is a bflat guest that runs every shim against 148 rows of precomputed results: 11 operand pairs for each non-shift op, and 10 shift counts on two values for each shift (counts include 0, 63, 64, 255, 256, 300 and 2^256-1). Each row runs three ways: into a separate buffer, with the result over the first operand, and with the result over the second. The in-place forms are how a caller works on a stack slot, and they only hold if the instruction reads both operands before it writes. The public output is the mismatch count and the check count.
• build-release.yml:
  • link-guest also links u256.cs and uploads it with guest.
  • The new run-u256-vectors job runs it on the pinned nethermindeth/openvm-runner:v2.1.0-rv64 digest and fails unless the guest reports 0 mismatches over 444 checks.
• README: one sentence on the new check.

No library or package change, so no version bump.

The vectors pin the extension's semantics on the rv64 line, from extensions/bigint/circuit at OpenVM 56d5e4db:

• add, sub and mul wrap modulo 2^256;
• the shifts take their count from the low byte of the second operand, so 256 shifts by 0;
• slt and sltu write 1 or 0 as a 256-bit word.

Testing

Built locally with the bflat image Nethermind pins (bflat-riscv64-11:2ee8a26) against libopenvm.a from Nethermind.OpenVM.Runtime 1.0.0-preview.1, then run on the pinned runner:

public_values=0000000000000000bc0100000000000000000000000000000000000000000000

That is 0 mismatches over 444 checks. As a negative control, I flipped one expected limb of a mul row: the guest reported 2 mismatches (out of place and in place).

The fork has no Actions run for the new job, so CI on this PR is the first run of run-u256-vectors itself.

Remarks

🤖 AI agent (Claude Code / Opus 5.5) on behalf of @LukaszRozmej