Summary

Fix the 37 failing EEST state-test fixtures that the evm statetest CLI surfaces against fixtures/state_tests from execution-spec-tests v5.4.0. After this change ./build/bin/evm statetest --workers=12 fixtures/state_tests reports 63556 / 63556 passing.

Before this PR the CLI failed 37 subtests — these are real Erigon validation gaps that TestState was masking with a lenient wrapper that swallowed expectException mismatches. The failures fell into two buckets:

• Pre-fork tx-type rejection (28 fixtures) — typed transactions accepted on forks before the EIP that introduced them: EIP-2930 type-1 on pre-Berlin, EIP-1559 type-2 on pre-London, EIP-4844 type-3 on Shanghai.
• EIP-4844 blob-tx structural validation (9 fixtures, all in static/Cancun/stEIP4844_blobtransactions) — blob txs with to == nil (TYPE_3_TX_CONTRACT_CREATION), an empty blobVersionedHashes list (TYPE_3_TX_ZERO_BLOBS), or a versioned hash whose first byte ≠ 0x01 (TYPE_3_TX_INVALID_BLOB_VERSIONED_HASH).

Changes

execution/types/blob_tx.go — close the EIP-4844 validity gaps in BlobTx.AsMessage

Added three checks that the EIP-4844 spec defines as transaction validity rules but that previously only ran during block validation (misc.ValidateBlobs), bypassing the state-test code path and any other caller that goes through AsMessage without a block context:

• Reject tx.To == nil (contract creation is not permitted for blob txs) with ErrNilToFieldTx.
• Reject empty BlobVersionedHashes with new sentinel ErrBlobTxnEmptyBlobs.
• Reject any versioned hash whose first byte is not kzg.BlobCommitmentVersionKZG with new sentinel ErrBlobTxnInvalidVersionedHash.

The pre-existing if !rules.IsCancun check is moved to the top of the method (consistent with the other AsMessage implementations) so that pre-fork rejections happen before any other work.

execution/tests/testutil/state_test_util.go — route the state-test runner through production validation

RunNoVerify used to build the Message from JSON via toMessage and feed it straight to ApplyMessage, bypassing every consensus-level validation. The runner now:

1. Decodes post.txbytes via types.DecodeTransaction for every EEST fixture (all 63556 carry it) and calls tx.AsMessage(signer, baseFee, rules) — the same code path real block processing uses. Tx-type-vs-fork rejection and the new EIP-4844 checks fall out of this automatically.
2. Falls back to toMessage only for the six in-tree corner-case fixtures under execution/tests/test-corners/state/* that don't ship txbytes. None of those exercise pre-fork tx-type rejection, so bypassing AsMessage validation there is fine.
3. Builds BlockContext early (before AsMessage) so a single blockContext.Rules(config) provides the chain rules AsMessage needs, and the same context is reused for the EVM construction afterwards — no duplicated rules-construction code.
4. Computes a pre-state root once right after MakePreState via domains.ComputeCommitment. This is essentially free (the trie state is still warm from MakePreState's own flush) and lets every early-return path return a real root that satisfies the test framework's post.Root check for rejected-tx fixtures.
5. Early-returns on every error with the pre-state root. No more applyErr bookkeeping — each if err != nil { return statedb, preRoot, gasUsed, err } returns immediately, and the post-state finalize/commit/computeCommitment path runs only on a successful apply.
6. Skips the coinbase touch on rejected txs. On pre-EIP-158 forks (Frontier, Homestead) a stray AddBalance(coinbase, 0) creates an empty account that doesn't get pruned, which would break the post-state root the fixture expects for an unapplied tx. The touch now runs only after a successful ApplyMessage.

Why route the test runner through AsMessage

The original state-test runner duplicated nothing — it just skipped validation. Two consequences:

• Real Erigon validation gaps (the 9 EIP-4844 ones, fixed in blob_tx.go) were invisible to TestState because toMessage never invoked the checks.
• Any new validation added to AsMessage in the future would still be skipped by tests, defeating the purpose of state-test coverage.

Decoding txbytes and calling production AsMessage removes both problems with no new test-only validator. The only test-specific code that remains is toMessage, used solely for the six corner-case fixtures that pre-date txbytes support.

Verification

• ./build/bin/evm statetest --workers=12 fixtures/state_tests — 63556 / 63556 pass (was 63519 / 37).
• go test ./execution/tests/ -run "TestState$|TestStateCornerCases$" — pass.
• make lint — clean.
• GOGC=80 make test-all — 222 packages OK, 0 failures.

Performance

12-worker evm statetest over the full fixture set on a tmpfs (ramdisk):

Slightly faster (~2.1%) despite the added pre-state ComputeCommitment: the new early-return path skips a redundant FinalizeTx + CommitBlock + ComputeCommitment on rejected-tx fixtures (where no state changed anyway), and the pre-state commitment is essentially a cache walk since MakePreState just flushed the same trie.

Test plan

• ./build/bin/evm statetest --workers=12 fixtures/state_tests returns 0 failures against EEST v5.4.0 fixtures_develop.tar.gz.
• go test ./execution/tests/ -run "TestState$|TestStateCornerCases$" passes.
• make lint is clean.
• CI All tests workflow is green.