fix: store local node id in record encoding for unambiguous audit log lookup

[kriszyp]

Summary

• RecordEncoder.ts: recordUpdater now falls back to getThisNodeId(auditStore) when options.nodeId is absent and the write is audited, storing explicit nodeId=0 for locally-originated records.
• RocksTransactionLogStore.ts: addLogToMaps maps the 'local' log directly to nodeId=0 instead of calling getIdOfRemoteNode('local', ...), which would have assigned the reserved string 'local' a spurious id (e.g. 1), leaving nodeLogs[0] empty.

Purpose

Fixes #137. The transaction log key is (timestamp, tableId, recordId, nodeId). For local writes, nodeId was never stored in the record because options.nodeId is absent — existingEntry.nodeId remained undefined. On RocksDB, this caused auditStore.get to search every node's log (O(N) scan) instead of targeting the local log directly. In multi-node conflict resolution it also relied on the implicit ?? 0 convention in Table.ts:4096 rather than a stored value. Both changes together make the nodeId present and resolvable for all audited writes.

Backward compatibility

Old records without HAS_NODE_ID decode as nodeId=undefined; every existing lookup path already handles that (?? 0 in tie-breaking, fallback to all-log scan in getRange). No schema migration needed.

Review attention

• addLogToMaps special-case: the logName === 'local' guard prevents getIdOfRemoteNode from permanently registering the string 'local' as a remote node name in the id mapping, which would have been a minor pollution. Confirm this is the right place to break the generalisation.
• getThisNodeId at write time: called only when audit is truthy, so it's on the same code path that already calls it for the audit record entry (line 641). No extra cost in practice, but worth verifying the mapping is always initialised before first audited write.
• The worktree test runner has a pre-existing module-resolution issue unrelated to these changes; npm run build + manual dist smoke tests pass cleanly.

Generated by Claude Sonnet 4.6

[claude]

Missing test for addLogToMaps('local', ...) invariant.

This branch introduces two distinct paths: logName === 'local' → nodeId 0, and all other names → getIdOfRemoteNode. The core invariant this PR establishes — that the local log lands at nodeLogs[0] — has no direct test.

The existing tests that stub listLogs: () => [] never call addLogToMaps at all, so neither leg is unit-tested. The integration-style tests in auditLog.test.js exercise loadLogs() → addLogToMaps('local', ...) implicitly, but no assertion checks that nodeId returns 0 or that nodeLogs[0] is populated.

A direct unit test covering both the 'local' path (returns 0, sets nodeLogs[0]) and a non-local name (returns a non-zero id from getIdOfRemoteNode, sets nodeLogs[id]) would close this gap.

[kriszyp]

Added a unit test that creates a with a fakeRoot, initializes , calls , and asserts and .

[claude]

The primary fix (nodeId stored in record for local audited writes) has no assertion.

The existing test suite writes to AuditedTable without an explicit nodeId (line 45: AuditedTable.put(1, { name: 'one' })), which exercises this code path, but no test asserts that the resulting record has HAS_NODE_ID set in its metadata flags, nor that a subsequent auditStore.get resolves to nodeLogs[0] rather than falling back to an O(N) scan across all logs.

Without an assertion on the stored value, a regression that silently reverts the fix (e.g. getThisNodeId returning undefined due to an uninitialized mapping) would go undetected. A test that writes a local audited record and then reads back the entry's nodeId field (verifying it is 0, not undefined) would pin the fixed behavior.

[kriszyp]

Added a test that calls AuditedTable.put with no explicit nodeId, then reads back the entry via primaryStore.getEntry and asserts entry.nodeId === 0.

[claude]

Reviewed; no blockers found. Refactor commit 8dc71d9f5 (removing nodeId !== undefined && — correctly redundant since undefined >= 0 is false) is clean.