When a user reports generation slowing at long context, here are the first five measurements to take, ordered to isolate the bottleneck fastest:

1. **Time-to-first