fix(load-tests): Base fee reservation and funder affordability

[meyer9]

Summary

• Enforce funder max-cost affordability before funding
• Reserve Base transaction fees correctly (without unrelated L1 fee reserve)
• Add builder coverage for preloaded sender nonce chains

Stack

PR 4/5. Base: feat/load-tests-cumulative-gps-pacing (#4293).

1. feat(load-tests): open-loop generation, adaptive pacing, and B20 #4291 — open-loop core
2. fix(load-tests): funding and nonce robustness #4292 — funding/nonce robustness
3. feat(load-tests): pace capped runs by cumulative gas #4293 — cumulative GPS pacing
4. fix(load-tests): Base fee reservation and funder affordability #4294 — Base fee / funder affordability ← you are here
5. fix(load-tests): harden open-loop runtime and configuration #4295 — runtime hardening

Test plan

• cargo test -p base-load-tests --lib
• Builder nonce-chain coverage in base-builder-core
• Funding fails cleanly when funder cannot cover max cost

[github-actions]

queued_funder_transactions is accumulated with saturating_add(queued_nonces.len()) across every txpool endpoint. If multiple nodes gossip the same queued transactions, the same nonces are counted multiple times, which could inflate the total above zero even when a single node's queued set was successfully drained by the earlier drop_sender_transactions call.

Consider deduplicating nonces (e.g. collect into a HashSet) or taking the max across endpoints rather than the sum, similar to what highest_txpool_nonce already does.

[github-actions]

Removing the .max(priority_fee) floor means submission_max_fee can now return a value less than priority_fee. This produces an invalid EIP-1559 transaction (maxFeePerGas < maxPriorityFeePerGas).

The call sites touched by this PR correctly pre-clamp priority_fee with .min(max_gas_price), but at least two existing callers do not:

• load_runner.rs:262 (calibration path)
• pacing.rs:1172 (batch signing path)

With the new default max_gas_price of 0.01 gwei, if base_fee exceeds 0.1 gwei then priority_fee = base_fee/10 will exceed max_gas_price, and these call sites will build transactions with maxPriorityFeePerGas > maxFeePerGas.

Either add the .min(max_gas_price) clamp to all remaining callers, or keep the .max(priority_fee) floor here as a safety net.

[github-actions]

queued_funder_transactions is summed across all txpool endpoints with saturating_add. If multiple nodes gossip the same queued transactions, identical nonces are double-counted.

This matters because the check at line 153 hard-errors when the total exceeds zero. A single queued transaction visible on two nodes would report a count of 2 even if drop_sender_transactions successfully cleared one of them — the second node may just not have propagated the removal yet.

Consider deduplicating nonces (e.g. collect into a HashSet<u64>) or taking the max count across endpoints rather than the sum, similar to the highest_txpool_nonce logic.

[github-actions]

Review Summary

Changes reviewed: Enforce funder max-cost affordability, clamp priority_fee to max_gas_price, separate pending/queued txpool nonces, add builder nonce-chain test.

Findings

1. submission_max_fee floor removal creates invalid EIP-1559 txs at unclamped call sites (correctness)
submission.rs:515 — The .max(priority_fee) floor was removed from submission_max_fee, but two existing callers (load_runner.rs:262, pacing.rs:1172) do not pre-clamp priority_fee with .min(max_gas_price). With the new 0.01 gwei default cap, any base_fee > 0.1 gwei will produce maxPriorityFeePerGas > maxFeePerGas, which is an invalid EIP-1559 transaction.

2. Queued funder transaction count inflated by multi-node double-counting (correctness)
funding.rs:138-139 — queued_funder_transactions is summed across all txpool endpoints. If the same queued nonce is visible on multiple nodes (normal gossip), the count inflates and the hard-error at line 153 may fire spuriously. Should deduplicate nonces or take max per-endpoint, consistent with how highest_txpool_nonce works.

Other notes

• The sender_transaction_nonces refactor to return (pending, queued) tuple is clean.
• The builder test (test_preloaded_nonce_chain_available_after_pruning_head) is well-structured and covers a meaningful scenario.
• Import hoisting in best_txs.rs test module follows project conventions correctly.

[github-actions]

Base Std historical fork tests

Fork	Result	Passed	Failed	Skipped	base/base	base-anvil	base-std
Beryl	pass	616	0	13	a9918116	6d744e03	4658f1b7
Cobalt	pass	709	0	14	a9918116	ae7557c4	8f88bf3e

View run