fix: retry PTY drain on empty select() to prevent data loss on macOS

The PTY drain logic in read_pty_output's finally block broke out of the
drain loop after a single empty select() poll (100ms). On macOS CI
runners under load (make test -j4 with concurrent package installs),
PTY kernel buffer delivery can be delayed beyond 100ms after subprocess
exit, causing the drain to miss the output entirely.

Replace the immediate break with a consecutive empty poll counter that
requires 3 empty select() calls (~300ms) before concluding the drain is
complete. This extends the tolerance window while still respecting the
existing DRAIN_TIMEOUT_SECONDS (2s) and MAX_DRAIN_BYTES limits.

Isolated reproduction confirms:
- Original: 0/20 pass rate when data arrives >100ms after drain starts
- Fixed: 20/20 pass rate for delays up to 250ms+

Fixes #821

Author: mangelajo

.agentready/assessment-20260525-094133.json

@@ -0,0 +1 @@
+assessment-20260525-094133.json