perf(evm): single-pass execute and reserve-based simulate gas limit

[kyscott18]

Depends on #57. This branch includes that commit. Merge #57 first.

Summary

Execute

• execute runs once, keeps that post-state, and records the touched accounts and slots as the journal's access list.
• execute's gas_used is the gas with cold access. No runtime code reads it.

Simulate

• simulate runs the discovery pass (skipped when the journals give a cached access list), then one pass at max gas with the access list warm. The binary search and the final pass are removed.
• gas_limit = gas_used + reserve. Each frame reserves 2,301 gas for the SSTORE sentry (EIP-2200), and enough that 63/64 of its remaining gas covers each child's gas plus the child's reserve (EIP-150). Precompiles get no SSTORE reserve. Monad has no refunds, so gas_used is the gas spent.
• At that limit, the tx takes the same path as at max gas. The limit can be up to about 2,300 gas above the true minimum.
• Behavior change: if a call halts (for example, a precompile runs out of gas or rejects its input), simulate returns an error. A halted call uses all the gas it gets, so no gas limit that the chain accepts can run the tx.

Other

• The access list leaves out all precompiles of the active spec. Before, the filter covered only 0x01–0x0a, so P-256 at 0x100 was in the access list.
• ExecuteResult and SimulateResult are separate types in Rust and TS. SimulateResult no longer has journal_id or slot_writes.
• harness.rs is simpler: one Journal type writes pre or post images, access lists merge through ordered sets, the tx fields are parsed once per request, and dispatch uses one response helper. The wire format does not change.
• The typewriter-evm test script now also runs cargo test --lib.

Results

Same machine, PR head before these changes (9008e45) against 4bc8675:

	before	after
packages/typewriter/test/harness.benchmark.ts (3 runs)	3,765–3,775 mutations/s	3,694–3,847 mutations/s
typewriter-evm simulate, discovered, burn 5000	5.15 ms	3.30 ms
typewriter-evm simulate, cached, burn 5000	3.72 ms	1.43 ms

The harness benchmark shows no measurable change: simulate does not limit it. typewriter-evm execute throughput varied from 99k to 133k calls/s between runs, so it gives no clear result.

Test plan

• bun run lint, bun run typecheck, bun run test
• Rust tests: the tx fails at gas_used and succeeds at gas_limit for an SSTORE sentry, precompile calls (0x01, 0x100), and nested calls. A halted call returns an error.
• The access list leaves out 0x100 and matches eth_createAccessList.

🤖 Generated with Claude Code

Confidence Score: 4/5

Fix the outstanding rejection of successful transactions with caught child failures before merging.

Findings

1.  Caught calls block submission ▶
2.  Gas tests lose coverage ▶

Fix with agent prompt

### Issue 1
packages/typewriter-evm/src/harness.rs:435-436
`frame_end` marks every halted child as impossible to run at any gas limit. A contract can call a child with a fixed gas budget, catch its `INVALID` or out-of-gas failure, and still succeed. The child consumes only that budget, but `finish` returns an error. This makes `simulate` reject the successful transaction and stops submission.

Track the child's gas budget instead of rejecting every halt.

### Issue 2
packages/typewriter-evm/src/harness.rs:897-900
The gas tracker was replaced, but the tests for `CREATE` and calls that send value were deleted without replacements. The remaining call tests all send zero value, so mistakes in those gas rules could pass unnoticed. Keep tests that run these transactions at the returned `gas_limit` and check that they succeed. This is non-blocking test feedback.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

execute now keeps the state from one cold-access run and saves the touched accounts and slots in its journal.

• Executed transactions run once and keep the cold-access result.
• Simulations estimate gas from one warm run and a reserve.
• Discovered access lists skip precompiles that are already warm.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  E[execute] --> D[Run without an access list]
  D --> A[Collect touched accounts and slots]
  A --> J[Save journal and commit successful state]
  S[simulate] --> U[Undo selected journals]
  U --> C{Cached access list?}
  C -->|No| F[Discover access list and discard state]
  C -->|Yes| W[Run with warm access list]
  F --> W
  W --> G[Return gas used plus frame reserve]
  G --> R[Discard simulated state and replay journals]

Loading

_{Reviews (3) · Last reviewed commit: "refactor(evm): replace the gas tracker w..." · Reviewed by Greptile}

[greptile-apps]

Caught calls block submission

frame_end marks every halted child as impossible to run at any gas limit. A contract can call a child with a fixed gas budget, catch its INVALID or out-of-gas failure, and still succeed. The child consumes only that budget, but finish returns an error. This makes simulate reject the successful transaction and stops submission.

Track the child's gas budget instead of rejecting every halt.

Prompt To Fix With AI

This is a comment left during a code review.
Path: packages/typewriter-evm/src/harness.rs
Line: 435-436

Comment:
**Caught calls block submission**

`frame_end` marks every halted child as impossible to run at any gas limit. A contract can call a child with a fixed gas budget, catch its `INVALID` or out-of-gas failure, and still succeed. The child consumes only that budget, but `finish` returns an error. This makes `simulate` reject the successful transaction and stops submission.

Track the child's gas budget instead of rejecting every halt.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Gas tests lose coverage

The gas tracker was replaced, but the tests for CREATE and calls that send value were deleted without replacements. The remaining call tests all send zero value, so mistakes in those gas rules could pass unnoticed. Keep tests that run these transactions at the returned gas_limit and check that they succeed. This is non-blocking test feedback.

Prompt To Fix With AI

This is a comment left during a code review.
Path: packages/typewriter-evm/src/harness.rs
Line: 897-900

Comment:
**Gas tests lose coverage**

The gas tracker was replaced, but the tests for `CREATE` and calls that send value were deleted without replacements. The remaining call tests all send zero value, so mistakes in those gas rules could pass unnoticed. Keep tests that run these transactions at the returned `gas_limit` and check that they succeed. This is non-blocking test feedback.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!