docs(coral-tpu): update README with ai-edge-litert, dynamic models, and watchdog

- Replaced references to PyCoral framework with ai-edge-litert
- Documented the new dynamic model download behavior at deploy-time (no longer tracked in git)
- Added a section on Resilience & Auto-Recovery covering the TPU Watchdog's hang (10s invoke timeout) and stall (30-frame silent drop) detection logic
- Refined Linux/macOS Python documentation since the constraint has been lifted to 3.9-3.13

Author: solderzzc

skills/detection/yolo-detection-2026-coral-tpu/README.md

@@ -16,8 +16,9 @@ flowchart TD
 
     subgraph Native_Process[Python Virtual Env]
         D[detect.py Script]
-        E(PyCoral Framework)
+        E(ai-edge-litert)
         F[yolo26n_edgetpu.tflite]
+        W[TPU Watchdog]
     end
     
     subgraph Hardware
@@ -28,26 +29,35 @@ flowchart TD
     Aegis -->|stdin JSONL 'frame' event| Native_Process
     D -.->|Loads| F
     C -.->|Reads JPEG| D
+    D <-->|Monitors| W
     D -->|Hands off tensors| E
     E <==>|USB 3.0 Inferencing| G
     Native_Process -->|stdout JSONL 'detections'| Aegis
+    Native_Process -.->|stdout JSONL 'tpu_error'| Aegis
 ```
 
 ### Flow Breakdown
 
 1. **Deployment Phase**: 
-   The `deploy.sh` script executes during installation, hooking directly into your OS's package manager. It provisions essential C-libraries (`libusb`), native Google driver binaries (`libedgetpu`), and creates an isolated `python3 -m venv` sandbox securely.
+   The OS-specific deployment scripts (`deploy-linux.sh`, `deploy-macos.sh`, `deploy.bat`) provision essential C-libraries (`libusb`), native Google driver binaries (`libedgetpu`), and create an isolated `python3 -m venv` sandbox securely.
+   * **Dynamic Models:** Instead of tracking massive raw binary files in git, the deployment script dynamically downloads the `.tflite` compiled models directly from Google's master Edge TPU repository using `curl` / `Invoke-WebRequest`.
 2. **Inference Loop**:
    - The host system (Aegis-AI) continuously records frames and saves a snapshot to `/tmp/aegis_detection/` memory cache.
    - Using standard input (`stdin`), Aegis-AI sends a brief JSON control sequence instructing the Python watcher script (`detect.py`) to process the frame.
    - The Edge TPU fetches the tensor, performs native hardware execution using `libusb`, and instantly evaluates bounding box predictions without triggering CPU payload spikes.
    - Results are streamed synchronously over standard output (`stdout`) to Aegis-AI.
 
+## Resilience & Auto-Recovery
+
+The integration features a native **TPU Health Watchdog** built directly into `detect.py`:
+* **Hang Detection**: If the USB connection fails and the `ai-edge-litert` delegate locks up, the invocation thread will hard-timeout after 10 seconds. The skill emits a `invoke_timeout` error and exits, prompting Aegis-AI to autonomously restart the sidecar process.
+* **Stall Detection**: If the TPU thermally throttles and begins silently returning zero objects for 30 consecutive frames (despite functioning perfectly seconds prior), the watchdog emits a `stall` telemetry event back to Aegis-AI.
+
 ## Platform Differences
 
-* **Linux**: Inherits Python dependencies explicitly from system-level tools via `apt-get` (due to Debian/Ubuntu `externally-managed-environment` guards). Uses the official Coral `apt` repositories.
-* **macOS (Apple Silicon)**: Uses the `feranick/libedgetpu` community fork to allow Edge TPU execution on ARM64 chips without relying on Rosetta 2.
-* **Windows**: Relies entirely on PowerShell bootstrapping and UAC elevation to quietly execute Google's `install.bat`.
+* **Linux**: Inherits Python dependencies securely and provisions Google's `libedgetpu1-max` driver via `apt-get` on Debian/Ubuntu systems. Uses `ai-edge-litert` to support modern Python (3.9 - 3.13) without requiring legacy `pycoral` or `pyenv` locking.
+* **macOS**: Fully supports Apple Silicon natively. Uses the `feranick/libedgetpu` community fork to allow Edge TPU execution on ARM64 chips without relying on Rosetta 2.
+* **Windows**: Relies on PowerShell bootstrapping and transparent UAC elevation logs to quietly execute Google's `install.bat`. Fallbacks to community wheels gracefully.
 
 ## Configuration Options