[Privacy/Thermal] WorkloadsSessionHost silently consumes 90%+ GPU on Copilot+ PCs without user opt-in

## Bug Description

On a Lenovo ThinkPad with AMD Ryzen AI 9 HX PRO 370 (Copilot+ PC), **9 instances of `WorkloadsSessionHost.exe`** silently consume **90%+ of the Radeon 890M GPU** for ONNX Runtime compute workloads — without any user opt-in or notification. This caused the laptop to reach **87°C above the keyboard**, with keys physically painful to touch.

The user **does not use Windows Recall, Copilot, or any Windows AI features**. Recall was never configured or opted into. Despite this, the full AI workload stack was deployed and running continuously in the background.

## Privacy and Consent Concern

This is not just a performance bug — it is a **serious privacy and consent issue**:

1. **No opt-in**: The user never enabled Recall or any Windows AI feature. The AI workload infrastructure was deployed silently via Windows Update / OEM provisioning (`SignatureKind: Developer`, not Store).
2. **43 WindowsWorkload AppX packages** were installed without consent, including:
- `WindowsWorkload.Data.PhiSilica.Stx.1` (on-device LLM)
- `WindowsWorkload.OnnxRuntime.Stx.9.3` (ONNX inference engine)
- `WindowsWorkload.EP.AMD.VitisAI.1.8` (AMD GPU/NPU execution provider)
- `WindowsWorkload.ScrRegDetection` (screen region detection)
- `WindowsWorkload.TextRecognition` / `ImageSearch` / `SemanticText` / `LanguageModel` etc.
3. **`WSAIFabricSvc`** (Windows AI Fabric Service) was running and set to Automatic — orchestrating these workloads silently.
4. The `Recall` Windows optional feature was **Enabled** by default on this Copilot+ PC.
5. Although our security audit confirmed **no screenshots were captured** (ukg.db did not exist, CoreAIPlatform.00 was empty) and **no data was exfiltrated** (no outbound network connections from these processes), the fact that AI models capable of screen capture, OCR, and semantic analysis were loaded and running on the GPU without consent is deeply concerning.
6. **I work with sensitive client data daily** — confidential business documents, financial information, and proprietary systems. Having AI models with screen capture and text recognition capabilities silently running in the background, without my knowledge or consent, is a serious breach of trust. If these models had been actively capturing, my clients' data could have been exposed. This is unacceptable for any professional user handling sensitive information.

## Impact

- **Thermal**: 87°C sustained, keyboard too hot to touch, passive thermal throttling at 100%
- **GPU**: 90%+ compute utilisation on Radeon 890M by WorkloadsSessionHost (9 instances)
- **Power**: Significant battery drain and unnecessary power consumption
- **Privacy**: AI models for screen analysis, text recognition, and image search running without user knowledge or consent
- **User trust**: Users who explicitly choose not to use Recall/Copilot should not have AI inference infrastructure silently consuming their hardware

## Steps to Reproduce

1. Purchase/receive a Copilot+ PC (e.g. Lenovo ThinkPad with Ryzen AI 9 HX PRO 370)
2. Set up Windows 11 normally — do NOT enable Recall or Copilot
3. Use the PC normally for daily work
4. Open Task Manager → GPU tab
5. Observe 9 instances of `WorkloadsSessionHost.exe` consuming 90%+ GPU compute
6. Touch the keyboard area above the CPU/GPU — it will be dangerously hot

## Resolution Applied (user-side)

The following steps resolved the issue — **Microsoft should not require users to do this**:

1. Killed all 9 `WorkloadsSessionHost.exe` processes
2. Disabled Recall via Group Policy: `HKLM\SOFTWARE\Policies\Microsoft\Windows\WindowsAI\DisableAIDataAnalysis = 1`
3. Disabled `WSAIFabricSvc` service (stopped + set to Disabled)
4. Disabled Recall scheduled tasks under `\Microsoft\Windows\WindowsAI\Recall\`
5. Temperature dropped from 87°C → 70°C within 15 minutes

## Expected Behaviour

- AI workload infrastructure (WorkloadsSessionHost, ONNX Runtime, Phi Silica, VitisAI) should **NOT run** unless the user has explicitly opted in to Recall or another Windows AI feature
- `WSAIFabricSvc` should not auto-start on systems where Recall has not been configured
- The Recall Windows optional feature should not be Enabled by default — it should require explicit user action
- GPU compute resources should not be silently consumed for features the user has not activated

## Desktop

- **OS**: Windows 11 Pro 26200 (25H2)
- **Device**: Lenovo ThinkPad, AMD Ryzen AI 9 HX PRO 370
- **GPU**: AMD Radeon 890M Graphics (4 GB VRAM), Driver 32.0.22024.4008
- **RAM**: 56 GB
- **BIOS**: R2XET37W (1.17), November 2025
- **ONNX Runtime package**: WindowsWorkload.OnnxRuntime.Stx.9.3 v1.2601.1252.0
- **WorkloadsSessionHost**: WindowsWorkload.Manager.1 v1.2511.1196.0

## Related Issues

- #1806 — WorkloadsSessionHost occupying NPU on ARM64 (similar root cause, different symptom)

The report names WorkloadsSessionHost.exe, WSAIFabricSvc, Windows AI packages, and related issue #1806, but it identifies no repository file, test, or entry point. First determine whether this behavior is owned by onnxruntime-genai or by Windows/OEM components, then compare the related issue and reproduce the workload behavior. Done requires a confirmed fix preventing unopted workloads from consuming GPU resources, or a clear ownership determination.

Written by the indexing model from the issue text.