Code Review Summary

Regression Check

All modifications from main have been verified:

1. Removed handleStreamClosed goroutine (write_stream.go) — Actually an improvement. The old code had two goroutines: handleResponses exited on error without failing pending futures, leaving that to handleStreamClosed (which used generic io.EOF). The new code consolidates both: handleResponses now fails all pending futures with the actual error on Recv() failure, which is better — CodeNodeIsNotLeader errors propagate correctly for hint extraction.

2. writeStream invalidation on leader hint (executor.go) — The RLock→Lock upgrade race is the same pattern as main (for failed streams). Worst case: two goroutines create streams simultaneously, one gets overwritten. Old stream is now explicitly closed and marked failed, which is better than main's silent overwrite.

3. shard_manager.go Leader fallback — Changed from panic("shard not found") to return s.serviceAddress. Safer fallback — the service address can redirect, whereas a panic crashes the client.

4. public_rpc_server.go error handling — Logic properly inverted. CodeNodeIsNotLeader now gets enriched with hint; all other errors still get logged as warnings. Behavior preserved and enhanced.

5. Batch retry paths (read_batch.go, write_batch.go) — hint variable is updated in RetryNotify callback (fires between retries), correctly set before the next execute call.

6. assignment_dispatcher.go — Index rebuilt atomically under write lock, queried under read lock. Initialized() fixed to use RLock.

Verdict

No regressions, no blocking issues. The PR is ready to merge.