db/state/execctx: fix AccountsDomain codeHash decode (DeserialiseV3) — correctness

[mh0lt]

Correctness fix

codeHashForAddr resolves an account's codeHash from the AccountsDomain so the CodeDomain ethHash bypass can serve shared bytecode (proxies, clones, ERC-20 holder sets) without an addr-keyed file read.

decodeAccountCodeHash decoded the account value with acc.DecodeForStorage, but AccountsDomain values are SerialiseV3-encoded. DecodeForStorage is the legacy MDBX bitmask format with an incompatible binary layout — applied to V3 bytes it silently misparses and leaves CodeHash empty.

Consequently codeHashForAddr returned nil for every account, and the ethHash bypass never engaged for any contract: every CodeDomain read that missed the addr cache fell through to a file read. This is a decoding-correctness bug — the wrong decoder applied to the stored encoding.

Fix

Use accounts.DeserialiseV3, the matching decoder for AccountsDomain values. One-function change; no behaviour change for valid encodings other than producing the correct codeHash.

Verification (instrumented mainnet-tip run)

The codeHashForAddr resolve path and ethHash bypass were instrumented:

metric	before	after
ethHash bypass served% (cache, no file read)	0	56 → 72 → 83% (climbs as dedup warms)
resolve produced no codeHash for a contract-with-code	~99%	0
GetByEthHash calls	0 (never reached)	active, 83% hit

Routing

The buggy decodeAccountCodeHash/codeHashForAddr is introduced by #21386 (StateCache LRU) and does not exist on main, so the fix targets #21386 — the appropriate point in the perf stack. It flows up to #21414 / #21416 / #20893 (all stacked on this branch). To be cherry-picked into those branches as needed.

🤖 Generated with Claude Code

[mh0lt]

Folding this fix directly into #21386 instead — the buggy code is introduced there (not on main), so updating the PR branch is cleaner than stacking a separate PR. Same commit (fd74979) pushed to mh/perf-statecache-lru-pr.