Files
precompile/blake3/accel_parity_test.go
zeekay ada296185c fix(accel): eliminate CPU/GPU verdict divergence in verify precompiles (consensus-split + forgery)
Under the enable-everything precompile policy these verify-only precompiles are
no longer refused under strict-PQ, so any divergence between the luxfi/accel
GPU/SIMD path and the pure-CPU path is a LIVE consensus split: GPU-equipped and
CPU-only validators would return different verify verdicts on the same input and
the chain forks. Several accel paths were also outright wrong-primitive forgery
surfaces.

LAW applied: the GPU verdict must be byte-identical to the CPU verdict on every
input, or the accel path is a consensus hazard and is removed. CPU (pure-Go,
CGO_ENABLED=0) is now the single source of truth for every on-chain verdict.
None of these GPU kernels can be proven byte-identical here, and several are
provably NOT, so every verdict-trusting accel branch is deleted and routed to
the canonical CPU verifier. Determinism over speed.

Active forgery/wrong-output in the DEFAULT build (accelcrypto CPU fallback
leaked non-CPU semantics, err==nil):
- sr25519: verified via the accel Ed25519 kernel -> a valid Ed25519 signature
  was ACCEPTED as a valid sr25519/Schnorrkel signature (wrong-curve forgery).
- blake3: accelcrypto.Hash(HashBlake3) returns BLAKE2b, not BLAKE3; the "real
  blake3" fallback was dead code. OpHash256 + MerkleRoot emitted BLAKE2b.

Latent splits (reachable only under -tags accel + a GPU backend):
- frost: FROST-Schnorr verified through the SigECDSA kernel (wrong primitive +
  32B x-only key mis-framed as a 33B ECDSA key).
- cggmp21: SigECDSA kernel truncated the 65B key to 33B and skipped the CPU
  path's recover-and-compare binding (ignored the v byte).
- ed25519: GPU Ed25519 kernel not proven byte-identical to the canonical
  verifier on cofactor / non-canonical-S / small-order edge cases.
- secp256r1: P-256 inputs sent to the fixed-curve secp256k1 ECDSA kernel.
- mldsa: GPU verified with EMPTY context; precompile binds the FIPS-204 ctx
  "lux-evm-precompile-mldsa-v1" -> different mu, different verdict.
- bls12381 / zk(groth16): GPU MSM result trusted without re-verification and
  with no curve binding carried to the kernel.
- kzg4844: GPU commitment used the backend's OWN embedded trusted setup (empty
  SRS) instead of the canonical EIP-4844 SRS -> breaks KZG binding.
- ai_mining (BatchVerifyMLDSA/BatchAccumulator): accel batch kernel truncates
  messages to 32 bytes and is mode-fixed -> diverges from VerifyMLDSA.
- fhe: GPU computed BFV arithmetic while CPU computes TFHE (different scheme,
  incompatible ciphertext, nil relin key) -> corrupt + divergent ciphertext.

Cleanups (no split, vestigial): removed dead mlkem encapsulateGPU; removed the
hpke GPU pre-warm that discarded its result and returned the CPU output anyway.
slhdsa was already CPU-only (prior identical fix).

Tests (differential / forgery proofs, all green CGO_ENABLED=0 and CGO -race):
- sr25519: a valid Ed25519 sig is now REJECTED; real sr25519 still verifies.
- blake3: OpHash256/MerkleRoot output == BLAKE3 and != BLAKE2b.
- cggmp21: flipped recovery byte is rejected (recover-compare authoritative).
- frost: a valid ECDSA sig is rejected as FROST-Schnorr.
- ed25519: verdict matches the canonical ed25519.Verify oracle over 64x4 vectors.
- mldsa: empty-ctx sig (the accel framing) rejected; precompileCtx sig accepted.

Net -593 LOC. No consensus code touched.
2026-06-27 19:46:11 -07:00

80 lines
2.6 KiB
Go

// Copyright (C) 2025, Lux Industries, Inc. All rights reserved.
// See the file LICENSE for licensing terms.
package blake3
import (
"testing"
refblake3 "github.com/luxfi/crypto/hash/blake3"
"github.com/luxfi/geth/common"
"github.com/stretchr/testify/require"
"golang.org/x/crypto/blake2b"
)
// blake3Ref computes the canonical BLAKE3-256 digest the precompile must emit.
func blake3Ref(data []byte) []byte {
h := refblake3.New()
h.Write(data)
out := make([]byte, DigestLength32)
h.Reader().Read(out)
return out
}
// blake2bRef computes BLAKE2b-256 -- the WRONG digest the removed accel
// "HashBlake3" path actually returned (accel/ops/crypto/crypto_cpu.go).
func blake2bRef(data []byte) []byte {
h, _ := blake2b.New256(nil)
h.Write(data)
return h.Sum(nil)
}
// TestBlake3_Hash256IsRealBlake3NotBlake2b is the regression proof for the
// wrong-hash accel divergence. accelcrypto.Hash(HashBlake3, ...) returned a
// BLAKE2b digest (its GPU branch has no BLAKE3 kernel and its CPU branch calls
// blake2b), and -- because accelcrypto.Hash returns a nil error on that CPU
// fallback -- the precompile's "real BLAKE3" branch was dead code in EVERY
// build. The OpHash256 output must now be BLAKE3 and must differ from BLAKE2b.
func TestBlake3_Hash256IsRealBlake3NotBlake2b(t *testing.T) {
cases := [][]byte{
{},
[]byte("abc"),
[]byte("the quick brown fox jumps over the lazy dog"),
make([]byte, 1000),
}
for _, data := range cases {
input := append([]byte{OpHash256}, data...)
gas := Blake3Precompile.RequiredGas(input)
out, _, err := Blake3Precompile.Run(
nil, common.Address{}, ContractAddress, input, gas, true,
)
require.NoError(t, err)
want := blake3Ref(data)
require.Equal(t, want, out,
"OpHash256 must output canonical BLAKE3 for len=%d", len(data))
require.NotEqual(t, blake2bRef(data), out,
"OpHash256 must NOT output BLAKE2b (the accel divergence) for len=%d", len(data))
}
}
// TestBlake3_MerkleRootUsesBlake3 proves the Merkle path is BLAKE3 too: the
// removed accel branch would have built the tree from BLAKE2b node hashes.
func TestBlake3_MerkleRootUsesBlake3(t *testing.T) {
p := &blake3Precompile{}
leafA := make([]byte, DigestLength32)
leafB := make([]byte, DigestLength32)
for i := range leafA {
leafA[i] = byte(i)
leafB[i] = byte(255 - i)
}
got := p.computeMerkleRoot([][]byte{leafA, leafB})
// Reference parent = BLAKE3(leafA || leafB).
want := blake3Ref(append(append([]byte{}, leafA...), leafB...))
require.Equal(t, want, got, "merkle parent must be BLAKE3(left||right)")
require.NotEqual(t, blake2bRef(append(append([]byte{}, leafA...), leafB...)), got,
"merkle parent must NOT be BLAKE2b")
}