commitment: preserve account load flags across storage folds

[taratorio]

Cherry-pick of #22603 onto main, without its prerequisite PR chain.

Preserves account load flags when folding storage subtrees in the hashed Patricia trie, preventing BAL-provided account values from being overwritten by stale state-reader values during ModeUpdate.

Validation

• go test ./execution/commitment -run '^TestModeUpdatePreservesAccountAcrossStorageFold$' -count=1
• make lint (two consecutive clean runs)
• make erigon integration

[Copilot]

Pull request overview

This PR cherry-picks a fix in the execution/commitment hashed Hex-Patricia trie folding logic to preserve account “loaded” flags when folding storage subtrees. This prevents ModeUpdate from re-loading (and potentially overwriting) BAL-provided account fields with stale state-reader values during storage collapses, and adds a regression test to lock in the behavior.

Changes:

• Preserve existing account-load flags when folding nodes in the storage portion of the hashed keyspace (fillFromLowerCell).
• Add a regression test that reproduces the “account updated + storage fold” pattern and asserts ModeUpdate matches ModeDirect.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
execution/commitment/hex_patricia_hashed.go	Adjusts fold propagation to merge (rather than overwrite) load flags when folding storage-depth nodes.
execution/commitment/hex_patricia_hashed_test.go	Adds a regression test ensuring ModeUpdate preserves account data across a storage fold and matches ModeDirect root output.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[mh0lt]

Approve — and empirically confirmed against a real consensus failure.

This fixes a pre-existing wrong-trie-root bug on the --experimental.bal (BAL-driven parallel commitment) path that I had independently root-caused on glamsterdam-devnet-6 at block 70097 (BAL-driven root 248a318a… vs correct e4e83f65…; reproduced on both main and #22409). My instrumentation pinned the divergence to the ModeUpdate fold: fold(197 updates)→248a vs incremental(227)→e4e8 on the same computeIsolated path — and, tellingly, re-emitting the "missing" account leaves at the changeset level was a no-op, which pointed at exactly this: the corruption is the trie-compute loaded flag, not a missing update. When a storage row (lowDepth > 64) folds up into an account cell, the old cell.loaded = lowCell.loaded clobbered cellLoadAccount, so the account looked unloaded and its BAL-provided value got overwritten by a stale state-reader read → wrong root. Merging the flags (addFlag) is correct and minimal.

Validation: built main + #22495 + this PR, ran with --experimental.bal + parallel commitment against a datadir at block 70096 on glamsterdam-devnet-6. It executed cleanly through 70097 and on to head 145030 with zero wrong-trie-root / BAL-mismatch events (previously it halted at 70097). So this resolves the bug end-to-end.

Context for #22409: it inherits this on its next merge-from-main; no conflict (does not touch hex_patricia_hashed.go). LGTM.