[kriszyp]

⊙ Problem

Replication receivers spend seconds per write on duplicates they already hold (#3467). The check looks up the write in the origin's transaction log. A base-copied record has no log entry, so on RocksDB the lookup reads to the end of the log. The duplicate is then logged again and re-sent to every peer, so an old stretch replays indefinitely.

🤔 Does this belong beside #3453?

Yes. #3453 hardens the deeper retry-path dedup. This PR covers first deliveries, which #3453 does not reach. Both edit the same design-note bullet, so whichever merges second keeps both guarantees.

💡 Solution

• A replicated write that matches the stored record's origin, version and log key, with no explicit refs, is skipped on first delivery as well as retry. It reads nothing from the log and appends nothing.
• The skip is not taken after an applied earlier write to the same key in the transaction. A publish or relocate updates the record's version without staging it.

⚖️ Alternatives

• Bound the exact-start lookup in rocksdb-js — still worth doing. A suffix-minimum index (rocksdb-js#676) would not bound this case: the re-appended segments hold keys from across the whole stretch.
• Purge an inactive origin's last log file — the stretch is re-written into new files, so retention never reaches it while duplicates keep being re-logged.

🤔 Should a base-copied record count as logged under its version?

I chose yes. A copy drops the record's head ref, so a distinct write from the same origin, whose version equals the copied write's log key, would be skipped. That needs a source-reported version equal to a commit timestamp. Carrying the head's log key in copy frames would close it.

🔧 Changes

• resources/Table.ts: the head-identity skip in commit drops the retry-only condition and requires priorWritesSkipped.
• resources/DESIGN.md: the existing note covers first deliveries and the publish/relocate exception.
• unitTests/resources/relayedHeadTieDuplicate.test.js: five cases through the replication subscription path.

✅ Verification

• Fails on base: a relayed duplicate of a base-copied record made 5 log reads and was appended to the log; a re-delivered two-write transaction read the log. Both pass with the fix.
• Guards pass on base and with the fix: a same-version write under a different log key is applied, and a put after a same-key message or relocate is applied. Both prior-write cases fail when priorWritesSkipped is removed.
• test:unit:resources: 4630 passing. Two failures (evictionBlobs resequenced-version case, replicatedDeleteRedelivery [put, delete] case) fail the same way on unmodified main.
• Not observable end-to-end without v4 migration history. The affected cluster can confirm on a 5.3 patch.

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

Release target: v5.3 — fixes code from #3366, a backport to v5.3; the affected cluster runs 5.3
Related PRs: #3453 overlaps (same design-note bullet), #3375 overlaps (changes relocate, which priorWritesSkipped relies on), 6 others independent
Complexity: complicated

_{Review-Coverage: authored=claude; ran=gemini,cursor-composer,codex,cursor-grok; adjudicated=domain; declined=muse; rounds=5; full=2 @ ab34002}

_{Review-Attention: study ~15m (critical: Table.ts; decisions: head-tie-on-first-delivery, skip-without-append, copy-row-log-identity, sibling-write-types; raised: degraded review) @ ab34002}

[github-actions]

Release cherry-pick v5.3: conflict

Cherry-pick onto v5.3 produced conflicts on commit(s): 6c10cca7992015539e9457d0c7df3139e5dce8e2 f8f32b712f2fbe5efe889b7951c0d6151c6139a5 2eb7630c5fa62c44bfe5008fed22a7db28091954

The conflict markers are committed on branch cherry-pick/v5.3/pr-3476.
A pull request has been opened to land this patch: #3478

[gemini-code-assist]

Code Review

This pull request updates the Harper design documentation and implementation to improve the handling of replicated duplicates. Specifically, it introduces a priorWritesSkipped helper function in Table.ts to ensure that replicated writes are correctly identified and processed, and adds comprehensive test coverage in relayedHeadTieDuplicate.test.js to verify these scenarios. There are no review comments to address for this pull request.