fix(simulator): bump pre-init buffer 1.05x -> 2.0x to fix CI

[meyer9]

Problem

The basic-benchmarks job in .github/workflows/public-benchmarks.yaml has been failing on every push to main since 2026-05-15 (PR #184). Symptom:

execution reverted: Not enough accounts to load/update

Reproduces deterministically on base-mainnet-simulation @ 25M gas (the first matrix cell, run against both reth and geth), about 7-8 benchmark blocks into a 20-block run. CI history:

Run	Date	Commit	Result
26758739814	2026-06-01	b5290a8 (#191)	❌ failure
26655072445	2026-05-29	4f98307 (#186)	❌ failure
26525697996	2026-05-27	(#188)	❌ failure
...			❌
25935745254	2026-05-15	cf47296 (#184)	❌ first failure
25870299556	2026-05-14	cb5b9d6	✅ last green

The contract revert is at Simulator.sol:82 — a benchmark-fidelity check that current_address_index + load + update <= num_address_initialized.

Mechanism

PR #184 (cf47296) changed targetCalls in the simulator worker:

-targetCalls := uint64(math.Ceil(float64(t.numCallsPerBlock) * t.scaleFactor))
+targetCalls := t.numCallsPerBlock

This was correct for scaleFactor > 1 (the 200M-gas case the PR was targeting). But for scaleFactor < 1 (the 25M case from base-mainnet-simulation, where scaleFactor = 25M / 35M = 0.714), the change INCREASES per-block consumption from ceil(46 × 0.714) = 33 calls/block to 46 calls/block.

Pre-init is sized as:

configForAllBlocks = base × (numCallsPerBlock × NumBlocks × scaleFactor × 1.05)

This gives a uniform 5% buffer post-PR-#184. The buffer is mathematically sufficient for an ideal run, but in practice the boundary is too tight: rounding accumulation across multiple Stats fields (per-tx blockCounts.Round(), big.Int.Int64() truncation, etc.) plus the deterministic per-tx layout pushes consumption past the safety margin well before block 20.

Fix

Bump the multiplier from 1.05 to 2.0. Pre-init now allocates ~2× the predicted consumption. Pre-init runs during setup as a one-time cost (dominated by mineAndConfirm RPC batching), so the wall-clock impact is a few seconds per cell — negligible against the ~5 min/cell, ~1 h/job total runtime.

-configForAllBlocks, err := t.payloadParams.Mul(float64(t.numCallsPerBlock) * float64(t.params.NumBlocks) * t.scaleFactor * 1.05).ToConfig()
+// 2.0x safety multiplier (was 1.05). ...
+configForAllBlocks, err := t.payloadParams.Mul(float64(t.numCallsPerBlock) * float64(t.params.NumBlocks) * t.scaleFactor * 2.0).ToConfig()

Verification

• go build ./runner/... clean
• go vet ./runner/payload/simulator/... clean
• go test ./runner/payload/simulator/... -timeout 60s clean (existing mineAndConfirmBatchSize tests pass; no test covers testForBlocks pre-init sizing directly)
• Mechanism is symmetric: only the SIZE of the pre-init batch changes, not what it does. Setup still runs InitializeAddressChunk / InitializeStorageChunk exactly the same way.

Out of scope (follow-up)

While digging I noticed two real bugs worth separate PRs:

1. OpcodeStats.Sub/Add iterate other rather than s ∪ other (types.go:48,38) → per-tx blockCounts.Precompiles is always empty in sendTxs → per-tx actual gas is much lower than the gas value calcNumCalls uses → block utilization is far below the configured gas limit even when the require passes.
2. Self-recalibrating numCallsPerBlock based on observed on-chain consumption would be more principled than the static safety multiplier and would also fix the base-mainnet-simulation @ 250M under-fill I documented in the gas-limit investigation. Likely a one-method addition to simulatorPayloadWorker.

Both are outside the scope of "make CI green again", which this PR does.

Draft because end-to-end validation requires one CI run.

[cb-heimdall]

🟡 Heimdall Review Status

Requirement	Status	More Info
Reviews	🟡 0/1	Denominator calculation Show calculation 1 if user is bot 0 1 if user is external 0 2 if repo is sensitive 0 From .codeflow.yml 1 Additional review requirements Show calculation Max 0 0 From CODEOWNERS 0 Global minimum 0 Max 1 1 1 if commit is unverified 0 Sum 1

[meyer9]

Superseded by #194 which combines this buffer bump with the OpcodeStats Add/Sub fix and observed-gas recalibration into one coordinated change.