Troubleshooting matrix

Symptom, likely cause, and fix in one place.

Diverges immediately → wrong measurement convention, stale response, or under-resolved readout → validate the integration with a bounded dry run.

Moves, then stalls → statistic noise or saturation → increase the batch slightly and tighten the operating range.

Works at small n, fails at scale → scalar bottleneck or stale corrections → preserve the per-channel measurement vector.

Hosted loop is too slow → network hop dominates → scope the offline build.