Problem
The Supervisor samples the Windows DXGI process budget and NVIDIA device memory, but it does not explain when those two views materially disagree. Under WDDM, a generous per-process budget can coexist with little genuinely free device memory.
Required insight
Add one insight object to the existing /api/insights computation and render that same object in the dashboard. It should:
- compare the sampled DXGI Budget with NVIDIA device used/free memory over the same window;
- explain material divergence in operator language;
- retain raw unsmoothed samples as evidence;
- state the measurement window and adapter identity;
- return unavailable when either source cannot be measured.
Do not infer another process's exact allocation from DXGI CurrentUsage; the supported conclusion is limited to the disagreement between the process budget and device-wide availability.
Acceptance
A fixture with a high DXGI budget and low device-free memory produces a measured divergence insight with raw figures. Missing either source produces unavailable rather than a clean result.
Split from #12.
Problem
The Supervisor samples the Windows DXGI process budget and NVIDIA device memory, but it does not explain when those two views materially disagree. Under WDDM, a generous per-process budget can coexist with little genuinely free device memory.
Required insight
Add one insight object to the existing /api/insights computation and render that same object in the dashboard. It should:
Do not infer another process's exact allocation from DXGI CurrentUsage; the supported conclusion is limited to the disagreement between the process budget and device-wide availability.
Acceptance
A fixture with a high DXGI budget and low device-free memory produces a measured divergence insight with raw figures. Missing either source produces unavailable rather than a clean result.
Split from #12.