Fixes #21651 — non-deterministic parallel-exec BAL computation rejecting canonical blocks as invalid and wedging the CL.

Root cause

BAL address-access recording is started by ibs.Prepare (allocates a fresh addressAccess map, sets recordAccess = true) and ended by the ibs.AccessedAddresses() harvest at the end of TxTask.Execute. The bug is that only regular user transactions call Prepare (reached via ApplyMessage → execution/protocol/txn_executor.go:423/590), while system tasks don't — yet system tasks still perform the harvest:

• block-init (TxIndex == -1) runs engine.Initialize + syscalls via SysCallContract — no ApplyMessage, no Prepare;
• block-end (IsBlockEnd()) executes no EVM code at all.

Both end in TxTask.Execute's success-path harvest (result.AccessedAddresses = ibs.AccessedAddresses(), execution/exec/txtask.go:624). On a clean worker IBS that harvest returns nil — recordAccess is false, so even the Initialize syscalls record nothing, which is the spec-correct behavior (system-call touches don't belong in the BAL). The system tasks therefore implicitly relied on the worker IBS being pristine.

That assumption broke in combination with two other facts:

1. A conflict-aborted incarnation never reaches the harvest — TxTask.Execute skips it when result.Err != nil (execution/exec/txtask.go:618) — so the aborted tx's touched-address set and recordAccess = true survive on the worker's long-lived IBS. IntraBlockState.Reset(), called before every task, cleared all other per-task state (versioned reads/writes, journal, EIP-2929 access list, transient storage) but not addressAccess/recordAccess. Regular txs were still immune because their Prepare overwrites the residue; system tasks were not.
2. Workers are shared by all in-flight block executors of a batch (single pe.in queue), so the system task that lands on the residue-laden worker usually belongs to a different block than the aborted tx. It scoops the residue as its own accesses, and nextResult records them into its block's blockIO (execution/stagedsync/exec3_parallel.go:2301).

AsBlockAccessList then creates an entry for every accessed address (execution/state/versionedio.go:1417) → phantom empty, address-only entries in the computed BAL. Which worker holds residue and which block's system task lands on it is schedule-dependent → the computed BAL hash differs across re-execution attempts of the same block. This explains all field symptoms: extra empty accounts, precompile-range extras, extras belonging to neighboring blocks' canonical BALs, and computed BALs with more accounts than the stored sidecar.

Reproduction (before the fix)

Local glamsterdam-devnet-5, erigon @ 1ca634d4 (ERIGON_EXEC3_PARALLEL=true) + prysm glamsterdam-devnet-5, full resync from genesis. During catch-up batch execution the node rejected canonical block 6638:

• computed BAL: 340 accounts; canonical: 339
• the extra entry 0xd4093C4A57D6F952849D6bf47e9a1F6CDCa79b7a is address-only (no reads, no changes), absent from 6638's canonical BAL, and legitimately active in a dozen neighboring blocks of the same batch (6600–6619)
• the post-unwind re-execution of 6638 computed 339 accounts and passed — non-determinism confirmed on the same block within one run

Same signature in 9 earlier mismatch dumps (blocks 1405–1430): every extra address was an empty entry appearing canonically in other blocks of the same batch.

Fix

Clear recordAccess/addressAccess in Reset() — i.e. make Reset() guarantee for every task what Prepare only guaranteed for user transactions, so a task's AccessedAddresses can only contain addresses touched during that task's own execution. This also covers any future Prepare-less task type, not just block-init/block-end. Audited the full IntraBlockState struct: this was the only accumulating field Reset() missed; no caller relies on recording state surviving Reset().

Testing

• TDD: TestAddressAccessResetInIBSReset written first — red on old code, green with the fix.
• ./execution/state/... (incl. -race) and ./execution/exec/... pass.
• All engine-api BAL tests pass, incl. TestEngineApiBALParallelConsistencyStress ×50 with -race (the canary for this bug class).
• Devnet validation: full glamsterdam-devnet-5 resync from genesis with the fixed binary — genesis → tip 11431 (delta 0 vs network), 0 BAL mismatches, 0 invalid blocks, through every previously-failing range (960, 1405–1430, 5545–6320, 6638); at tip NewPayload/FCU return VALID and prysm imports live blocks normally.
• make lint clean (×2).