fix(txn): retry ERR_TRY_AGAIN on the same transaction (native in-place reset)

[kriszyp]

Pairs with HarperFast/rocksdb-js#710. rocksdb-js 2.5.0 is now published; rebased onto latest main (which had independently picked up the same dependency bump) and verified against the real native build — see Validation below.

Problem

#1696 fixed the #1695 source-apply ERR_TRY_AGAIN spin by replaying the writes onto a fresh RocksTransaction, because the native layer left the stranded snapshot in place so recommitting the same transaction never converged. But that workaround leaned on a rocksdb-js defect: the transaction-log publish gate (!IsBusy) published the change-feed entry on the failed commit, and the fresh replay (marked isRetry) relied on that premature publish to keep the entry from being lost — an entry visible ahead of its data for every reader in the window, and a permanent phantom if the retry was ultimately abandoned.

Fix

rocksdb-js#710 makes the native layer reset the transaction onto a fresh snapshot on a failed TryAgain commit — exactly as it always did for IsBusy — and publish log entries only on a real commit. So the fresh-transaction replay becomes both unnecessary and wrong (a fresh transaction with isRetry would never publish the now-unpublished entry, re-manifesting #1695). This PR:

• Recommits the SAME transaction on ERR_TRY_AGAIN, like ERR_BUSY always did. Its committedPosition survives the native reset (WAL write-once, rocksdb-js#668) and its onCommit hook stays attached, so the staged change-feed entry publishes exactly once, only when the retry really commits.
• Aborts the native transaction on coordinated-retry give-up (pre-existing gap surfaced by review: the RETRY_NOW exhaustion path threw without releasing the handle, leaving an unpublished log position pinning the committed-read watermark until GC; the rejection-path give-up already aborted).
• Poisons and aborts the whole linked chain on give-up, not just the head, via a shared abortChainAfterRetries() helper mirroring abortDueToTimeout()'s two-pass pattern — and guards each link's wrapper cleanup (txn.abort(), which can throw from savedBlobs/getEntry()) so one link's failure can't strand later links' native handles/read snapshots.
• Bumps @harperfast/rocksdb-js to ^2.5.0 — the paired-fix floor.

The Table.ts appendedAuditEntry guards from #1696 are deliberately kept: they already govern the ERR_BUSY same-transaction recommit, and the scenario they defend against is less reachable now (failed attempts never publish). Whether the TryAgain-specific rationale is now dead code deserves its own analysis — follow-up issue to come.

Validation

Against the real, published rocksdb-js 2.5.0 build:

• unitTests/resources/sourceApplyConflictRetry.test.js — including the real compact()-induced ERR_TRY_AGAIN repro, with the assertion inverted from fresh-transaction-id to same-transaction-id — plus a new focused unit test forcing one chain link's wrapper cleanup to throw and asserting later links are still detached/untracked/natively aborted. 5/5 passing.
• npm run test:unit:resources: 1249/1249 passing.
• npm ci, format:check, lint:required: clean.
• CI: Format Check, Lint, Integration Tests, Claude PR Review all green. Unit Test had one unrelated flaky failure (caching.test.js, timing-based) that passed on rerun. The advisory, non-blocking Next.js Integration Tests (downstream) check fails on a pre-existing main type error (unrelated to this diff, already being fixed on a separate branch).

Review history

Two review passes, both addressed:

Cross-model review (Codex ×2 + Harper-domain adjudication; Gemini leg failed — agy hang):

• Blocker (addressed): version coupling not encoded → ^2.5.0 floor.
• Verified & fixed: coordinated give-up abort (above).
• Adjudicated no-new-hazard: the appendedAuditEntry double-apply scenario requires a failed attempt's entry to be visible during retry — exactly what the new publish gate removes.

Live discussion review (Codex + Harper domain pass) after 2.5.0 published:

• Blocker (addressed): lockfile still resolved 2.4.0 after the manifest bump; regenerated against the real 2.5.0 publish and rebased onto main's independent bump of the same dependency to restore a clean merge (which had also silently stalled pull_request-triggered CI).
• Significant (addressed): abortChainAfterRetries()'s per-link wrapper txn.abort() was unguarded — a throw on an early link (from savedBlobs/getEntry()) exited the cleanup loop before later chain links were touched, contradicting the two-pass design's own guarantee (also independently flagged by Gemini earlier in review — see inline thread). Fixed by guarding that call too, with a regression test.

KrAIs, via Claude

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates @harperfast/rocksdb-js to version 2.5.0 and refactors transaction retry logic in DatabaseTransaction.ts so that both ERR_BUSY and ERR_TRY_AGAIN errors recommit the same native transaction, which is now reset in place onto a fresh snapshot. It also adds transaction aborting when retries are exhausted to prevent handle leaks. The reviewer recommends that when aborting a chain of linked transactions, all links in the chain should be poisoned and closed first before aborting each individual link to avoid leaking native handles for other transactions in the chain.

[claude]

Reviewed; no blockers found.

[cb1kenobi]

LGTM