[kriszyp]

⊙ Problem

This brings #3453 to v5.3. On RocksDB, a replicated write that retries after a commit conflict skips every audit duplicate check. If another delivery already committed, the retry applies it again and a CRDT add counts twice: under a newer write, beside an out-of-order sibling ref, or on the replay after an open read snapshot.

🤔 Merge to v5.3 before main?

Yes, as you asked. The port is #3453 as of 723f4ff, before main's #3510, which v5.3 does not have. #3453 itself now carries a merge of #3510.

💡 Solution

Ported with v5.3's single log key: main's receivedLogKey and appendedLogKey both become txnLogKey.

• On a retry, the walk skips a duplicate only when it reached the entry through links from committed records with known origins (check; ⚠️ look hardest here).
• A base restamped by an earlier publish or relocate in the same retry round does not count as committed.
• Keyed lookups stay off on a retry (keyed probe).
• The open-snapshot replay stops claiming an audit entry its abandoned handle never wrote.

🤔 Keep the numeric origin compare on v5.3?

Yes. #3453 dropped it when it merged #3510, where a per-handle minted key makes it redundant. v5.3 has no minted keys, so it still keeps two unmapped origins at one version apart (compare, test).

⚖️ Alternatives

• Fix it natively (#3253: check conflicts before the log flush, so no orphan exists) — the better alternative planning review named; it is a rocksdb-js change, and this port stays correct after it lands.
• Skip on an exact match reached through a link with no recorded origin — rejected by owner ruling: the link can lead to an orphan, which drops a write.

🔧 Changes

• resources/Table.ts: retryBaseIsCommitted seeds walkOriginKnown, carried through previousVersion links and sibling refs; the identity skip and equal-version tie need it on a retry. hasSavedSameKeySibling counts a same-key sibling only if it staged in the same handle and retry round. The stagedOwnAuditEntry comment holds the rationale.
• resources/DatabaseTransaction.ts: writes record the handle and retry round they staged in; the replay clears the audit flags for writes on a non-durable abandoned handle.
• resources/DESIGN.md: keeps v5.3's text; the duplicate-identity bullet gains the retry rules, and the durable-log note says an abort publishes a readable orphan.

✅ Verification

• replicatedRetryDuplicate.test.js is byte-identical to Stop retried replicated writes from re-applying a duplicate that already committed #3453 at 723f4ff (same-key batch, terminal-abort orphan). With v5.3's original Table.ts and DatabaseTransaction.ts, 6 of its cases fail; with this branch, all pass.
• replicatedRetryDuplicate + relayedHeadTieDuplicate: 45 passing. test:unit:resources: 4170 passing, 0 failing.
• The Review-Attention decision list also carries slugs from a first review round that ran against main by mistake; that round reviewed none of this diff.
• Review raised that N same-key writes in one retry check their siblings N² times. identityBase already loops the same chain; on a retry this calls it once more per write, the same order. N is the writes to one key in one transaction, normally 1–3.

Framing-Verdict: better-alternative-exists (b55f8f526ef5)

Refs #3361

_{🤖 Generated by Claude Code (Claude Opus 5.5); posted via @kriszyp.}

Release target: v5.3 — cherry-pick of #3453 before #3510

Related PRs: #3453 overlaps (source), #3478 overlaps (merged; the v5.3 base this extends), #3510 overlaps (main-only; why this ports 723f4ff), #3476 overlaps (merged on main; ported in #3478), #3499 overlaps (merged on main; ported in #3478), 9 others independent
Complexity: complicated

_{Review-Coverage: authored=codex; ran=cursor-composer,gemini; declined=claude,cursor-grok,muse,domain; rounds=2; full=1 @ 94d6143}

_{Review-Attention: deep ~470m (critical: DatabaseTransaction.ts, Table.ts; decisions: do-less-alternative, ambiguous-chain-policy, native-root-cause, ambiguous-retry-provenance, logged-commit-liveness, portable-bundle, restore-lock-fallback, named-process-lifetime, archive-capabilities; raised: degraded review, earlier round at this SHA) @ 94d6143}

[gemini-code-assist]

Code Review

This pull request refines the handling of replicated duplicates and retry behavior within transactions, specifically addressing how committed identity is determined before audit I/O. It introduces tracking of the specific transaction handle (committedOn) and retry round (committedRound) for each write operation. This ensures that sibling writes to the same key are only considered as the commit base if they were saved in the current retry round. Additionally, it improves audit walk validation on retries by requiring known committed-origin provenance (walkOriginKnown) to prevent treating orphaned, uncommitted audit entries as proof of a committed duplicate. Extensive unit tests have been added to verify these behaviors under various conflict and retry scenarios. I have no feedback to provide as there are no review comments.