Note

Low Risk
Low risk observability change: adds new Prometheus gauges and a small trainer-side hook to report packer buffer sizes/round-robin state, with no impact to training logic beyond metrics emission.

Overview
Improves multi-run packer observability by exporting per-run buffer backlog and the packer’s round-robin position as Prometheus metrics.

MultiPacker now exposes get_buffer_stats(), the training loop conditionally reports these stats to MetricsServer, and metrics_server.py adds/cleans up the new trainer_packer_buffer_length{run=...} and trainer_packer_round_robin_position gauges.

^{Written by Cursor Bugbot for commit 9701bfe. This will update automatically on new commits. Configure here.}