Summary

With the state cache enabled (--exec.state-cache=true) and parallel execution (EXEC3_PARALLEL=true, the default), block execution hits a deterministic invalid block — a transaction never converges in OCC validation:

[4/6 Execution] invalid block: could not apply tx <n>: too many validator-invalid retries: 251, expected: 250

The node then cannot make progress (every fork-choice update re-hits the same block and fails). Reproduces at the same block/transaction across clean restarts, so it is not timing-dependent.

Observations

• State cache off → executes clean. So the state cache is the trigger.
• Disabling the L2b code-by-hash bypass (SC_DEBUG_DISABLE_L2B=true) does not help — the failure recurs at the identical block/tx.
• Independent of the parallel changeset-generation fix (execution: fix parallel exec to produce changesets at end of each batch as serial #21659) — reproduces with and without it.

So a state-cache value is inconsistent with what the parallel executor's OCC validation expects, at this point in the chain. Root cause not yet pinned.

Introduced by the state-cache work in #21386.