fix(tests): execute the empty-account EXTCODEHASH hash assertion

[danceratopz]

🗒️ Description

test_extcodehash_empty_account asserts two things about EXTCODEHASH on an empty account: the access gas cost (no EIP-8038 code-read surcharge), and that the returned hash is 0. The gas assertion (storage slot 0) was always exercised, but the hash assertion (storage slot 1) was dead code and never ran.

The test concatenates the hash store after the gas measurement:

code = CodeGasMeasure(
    ...,
    sstore_key=gas_slot,
) + Op.SSTORE(hash_slot, Op.EXTCODEHASH(empty_addr))

CodeGasMeasure appends Op.STOP by default (stop=True), so the measure block halts before the trailing SSTORE. Slot 1 was therefore never written and kept the EVM default of 0. Because the expected value registered for slot 1 is also 0, the post-state check passed trivially: the test never observed the actual EXTCODEHASH return value. This is well hidden, since 0 is simultaneously the expected empty-account hash and the never-written storage default.

Fix

This PR passes stop=False to CodeGasMeasure so execution flows from the gas SSTORE (slot 0) into the trailing hash SSTORE (slot 1), then halts at the end of the code. Slot 1 now genuinely holds the opcode's return value, so the hash assertion has teeth: a client returning a non-zero hash for an empty account would now fail the test.

The hash store must run after the measured access, not before: in the cold case a pre-measure EXTCODEHASH(empty_addr) would warm 0xDEAD and turn the measured access warm, breaking the gas assertion. stop=False keeps the store after the measured region, where the account is already warm and only the return value matters.

This PR also adds a defensive sentinel so the assertion cannot silently regress. Before the measured region, slot 1 is poisoned with a non-zero value that the real store must overwrite back to 0. Because the empty-account hash 0 is indistinguishable from an unwritten slot, this is what makes the guard durable: if the trailing hash store is ever stranded again (for example by a future revert of stop=False), slot 1 keeps the sentinel and the test fails instead of passing vacuously. The poison is written before the first GAS reading, so it does not affect the measured gas.

Expected post-state values are unchanged. Slot 0 (gas) stays 0xbb8 (cold) / 0x64 (warm) and slot 1 stays 0; the only behavioral change is that slot 1 is now actually written. Execution traces confirm this: before the fix the trace shows a single SSTORE (slot 0) followed immediately by STOP. After the fix the trace shows the sentinel store to slot 1, then the unchanged gas store to slot 0, then the real hash store overwriting slot 1 back to 0. Reverting stop=False with the sentinel in place makes both the cold and warm cases fail, confirming the guard.

🔗 Related Issues or PRs

Fix-up to ethereum#3033 (EIP-8038 state-access gas-cost test suite), which introduced this test.

[marioevz]

LGTM.