feat: top-of-block pre-simulation to filter reverting tx spam

[dmarzzz]

Summary

• Adds a new PreSimulateOrders pipeline step that simulates pending pool transactions at the top of each block before the flashblock building loop, removing reverting txs from the pool so they never consume building time on the critical path.
• Uses a random coinbase during simulation (configurable) so adversaries can't detect the simulation environment via the COINBASE opcode — same anti-gaming technique as rbuilder's OrderSimulationPool.
• Step is placed after OptimismPrologue/FlashtestationsPrologue so it sees post-sequencer state. Bundles are deferred to the existing RemoveRevertedTransactions step. EVM-level errors (wrong nonce, etc.) are kept since they may become valid later.

Why

Builder time is being burned simulating adversarial reverting txs during flashblock production. RemoveRevertedTransactions currently filters them but only after the expensive simulation has already run on the critical path. This moves the filtering to the top of the block (once per payload job), so the flashblock loop sees a pre-cleaned pool.

This is the L2 analogue of rbuilder's two-phase approach on L1: pre-sim → filter → build, with random coinbase so tx.origin == known_builder style detection doesn't work.

Configuration

Two new CLI flags (both default true):

• --builder.presim (env PRESIM_ENABLED)
• --builder.presim-random-coinbase (env PRESIM_RANDOM_COINBASE)

Metrics

New metrics emitted by the step:

• txs_simulated_total, txs_dropped_total, gas_saved_total (counters)
• txs_simulated_per_job, txs_dropped_per_job, gas_saved_per_job (histograms)
• presim_duration (histogram, seconds)

Test plan

• cargo check — clean for new code (pre-existing flashtestations typos unrelated)
• cargo clippy — clean for new code
• cargo +nightly fmt --check — clean for new code
• cargo test --workspace --all-features — run full suite
• Manual: spin up against playground, send a reverting tx, confirm it's filtered before building (grep "presim: dropping reverting transaction" in logs) and metrics are emitted
• Manual: toggle --builder.presim-random-coinbase=false and confirm a tx that branches on COINBASE to revert can detect simulation (negative test for the anti-gaming property)
• Confirm flashblock build time histograms drop under the same adversarial workload that motivated this PR

Tests added (src/tests/presim.rs)

• presim_filters_reverting_tx_without_revert_protection — the key test: presim ON, revert protection OFF. Proves presim alone is sufficient.
• presim_and_revert_protection_both_active — belt and suspenders.
• presim_keeps_valid_transactions — no false positives.

Open questions / follow-ups

• Should there be a presim time budget (e.g. --builder.presim-time-budget=200ms) to bound the worst case if the pool is huge or sims are unusually expensive? Easy to add.
• Bundles are intentionally skipped here. Worth a separate pass to pre-sim bundles too once we've validated the loose-tx path.
• remove_any_with removes from both order pool and system pool. If a loose tx is also part of a bundle, the bundle gets dropped too. Unlikely in practice but worth flagging.

🤖 Generated with Claude Code

[dmarzzz]

Wrong repo — redoing this in op-rbuilder where the work is actually needed.