fix(tests): measure on-chain SSTORE regular gas and warm the warm rows

[danceratopz]

Stacked on ethereum#3033. This PR targets its eip-8038-tests branch and only changes tests/amsterdam/eip8038_state_access_gas_cost_increase/test_sstore_gas.py, fixing the test_sstore_regular_gas case that PR adds.

Why

test_sstore_regular_gas asserts the EIP-8038 regular SSTORE schedule, but it never measured the executed cost. It only cross-checked the framework gas model against a hand-derived formula (a closed tautology, since both come from the same gas_costs) and asserted post-state {slot: new}, which is independent of the gas charged. Separately, slot warmth came only as a side effect of a prep SSTORE emitted when current != original, so the four key_warm=True rows with current == original (00x_warm, xx0, xxy_warm, xxx) ran cold on-chain and produced fixtures identical to their cold siblings. The warm storage-access price was never exercised in execution.

What changed

• Wrap the measured SSTORE in CodeGasMeasure and assert its executed regular cost equals expected_regular. The model cross-check is kept as a secondary guard.
• Warm the key_warm rows via the transaction access list instead of relying on an incidental prep write.
• Leave gas_limit unset so the EIP-8037 state-gas reservoir lands above the EIP-7825 cap and Op.GAS measures regular gas only. The prior gas_limit=1_000_000 was below the cap, which zeroes the reservoir and spills the zero-to-nonzero rows' state gas (97920) into the measurement; with it unset the reservoir absorbs the state gas, matching test_create_regular_gas.
• Clarify the row ids. Every row now carries an explicit _cold / _warm / _dirty suffix plus a legend. _dirty (current != original) reuses the clean/dirty vocabulary already documented in tests/benchmark/stateful/bloatnet/test_single_opcode.py.

After the fix the warm and cold rows genuinely diverge. 00x_warm and xxy_warm measure 10100 against 13000 for their cold siblings, xxx_warm measures 100 against 3000, and xx0_warm is a real warm clear at 10100.

Verification

• uv run fill --fork Amsterdam tests/amsterdam/eip8038_state_access_gas_cost_increase/test_sstore_gas.py passes (30/30).
• just static is clean.

Per-row facts from a trace dump (--traces --evm-dump-dir). Across all rows measured reg equals stored@0 equals expected. The 97920 measured state on the 00x rows is charged on the state dimension and excluded from measured reg (no spill), and the _dirty rows show their unmeasured prep write.

id	prep	measured reg	measured state	stored@0	expected
00x_cold	none	13000	97920	13000	13000
00x_warm	none	10100	97920	10100	10100
0x0_dirty	10100	100	0	100	100
xx0_warm	none	10100	0	10100	10100
xxy_cold	none	13000	0	13000	13000
xxy_warm	none	10100	0	10100	10100
xyz_dirty	10100	100	0	100	100
xyx_dirty	10100	100	0	100	100
xxx_warm	none	100	0	100	100
xxx_cold	none	3000	0	3000	3000

🔗 Related Issues or PRs

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

[marioevz]

Two comments on these changes, thanks!

[marioevz]

This kinda defeats the purpose of gas_cost(fork)/regular_cost(fork)/state_cost(fork), and forces us to have to update this particular test if the gas price formula changes in the future.

We should simply:

expected_regular = measured_bare.regular_cost(fork)

and remove other gas calculations from this function.

[marioevz]

My comment should really be in the expected_regular = access_cost + write_cost line, but it also applies here that we should not assert the cost of the push opcodes via 2 * very_low either.

[marioevz]

Great idea, nice.