fix: drain buffered ACKs when errorsCh closes before acksCh (#258)

closes #233 

## Summary
- `readAcksLoop` defers `close(errorsCh)` before `close(acksCh)` (LIFO),
so `errorsCh` closes first.
- The `select` in `readAcks` could non-deterministically pick the closed
`errorsCh` and return, abandoning buffered ACKs in `acksCh` (buffer size
20).
- Lost ACKs leave successful appends in `inflightQueue`, which then time
out and get retried on a new session, causing data duplication.
- Fix: drain `acksCh` when `errorsCh` closes to process all pending ACKs
before returning.

## Test plan
- [ ] Verify `go build ./...` passes
- [ ] Confirm that under high-throughput pipelined appends, session
teardown does not lose buffered ACKs

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: infiniteregrets

s2/append_session.go

@@ -423,6 +423,11 @@ func (r *AppendSession) readAcks(session *transportAppendSession) {
 
 		case err, ok := <-session.errorsCh:
 			if !ok {
+				// errorsCh closes before acksCh (defer order), so
+				// drain any buffered ACKs before returning.
+				for ack := range session.acksCh {
+					r.handleAck(session, ack)
+				}
 				return
 			}
 			r.handleSessionError(session, err)