fix(replication): emit reload markers for user-DB tables so live subscribers recover copy-applied rows

[kriszyp]

Summary

copyApply covers all RocksDB copies, but its base-copy rows carry no per-row audit entries, so a live MQTT/SSE/WS subscriber on a freshly-copied node never sees them — they are durably stored but never delivered to an already-connected subscriber.

This is the harper-pro half of #495. Generalize emitSystemReloadMarkers → emitCopyReloadMarkers: once a copy is durable, emit a whole-table reload marker per table for any copied RocksDB database — the system DB keeps the fixed cluster-machinery list (SYSTEM_RELOAD_TABLES), a user DB emits one per table. The paired core change re-delivers each table's current scope to its live subscribers on that marker.

Stacked on #491 (base branch kris/copyapply-reload-marker) — it depends on emitSystemReloadMarkers/SYSTEM_RELOAD_TABLES introduced there. The core companion (harper#1491) is already merged to main, so the core fix targets main directly: HarperFast/harper#1530. This branch bumps the core submodule to that fix. Will rebase onto main once #491 lands.

Where to look

• replication/replicationConnection.ts — emitCopyReloadMarkers (was emitSystemReloadMarkers): for a user DB it iterates Object.keys(tables); the marker is LOCAL_ONLY, so the replication send path skips it (extendedType & LOCAL_ONLY → skipAuditRecord) and it is never forwarded to peers. Each node re-snapshots its own subscribers off its own copy-complete.
• integrationTests/cluster/aclConnectCopyReload.test.mjs — deterministic repro: add_node isLeader:true forces an audit-less base copy of pre-existing rows; a subscriber that is live before the copy must receive them. Verified as a true negative control (subscriber sees [] with the re-snapshot disabled).

Companion PR

• core (harper): fix(subscriptions): re-deliver current scope to live subscribers on a copyApply reload marker harper#1530 — re-delivers current scope on the reload marker (targets main; merge before/with this).

Verification

• New aclConnectCopyReload test passes; fails with the core re-snapshot disabled.
• Existing aclConnectCrossNode (3 tests) still green.

Cross-model review

Thorough (Codex + Gemini + Harper-domain adjudication): no blockers; all outside-model blockers were false positives on source trace. Domain pass confirmed LOCAL_ONLY non-replication and the system-DB gate preserving knownNodes.

Generated by an LLM (Claude Opus 4.8).

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request resolves an issue where live MQTT, SSE, or WS subscribers fail to receive records that arrive on a node via a copyApply base copy. Because these base-copy rows are written as snapshots without per-row audit entries, the live subscription notify path is bypassed. The PR addresses this by renaming emitSystemReloadMarkers to emitCopyReloadMarkers and expanding its scope to emit reload markers for all tables in user databases, prompting subscribers to re-read and recover the snapshotted rows. A regression test has also been added to verify this behavior. The reviewer's feedback suggests a valuable optimization to skip emitting reload markers for non-replicated tables, which avoids redundant write transactions and unnecessary re-reads.

[claude]

Reviewed; no blockers found.