## Task 1: CLI Flag Addition

- Decision: Used argparse instead of click to match explicit task requirement
- Decision: Preserved all existing CLI options with same defaults as pipeline.py
- Decision: Module entry point maintained: python -m opengait.demo still works


## Task 1 Fix (Retry)

- Decision: Use inspect.signature for forward compatibility
- Decision: Conditionally pass visualize kwarg only if constructor accepts it
- This allows Task 3 to add visualize parameter without breaking Task 1


## Task 2: OpenCVVisualizer Design Decisions

### Architecture
- Class-based design encapsulates state (mask_mode, windows_created)
- Lazy window creation via _ensure_windows() - windows created on first update()
- In-place drawing methods (_draw_bbox, _draw_text_overlay) avoid unnecessary copies

### Display Choices
- DISPLAY_HEIGHT=256, DISPLAY_WIDTH=176 (4x upscale from 64x44 silhouette)
- INTER_NEAREST interpolation preserves pixelated look of silhouette
- Side-by-side view (mode 0) converts to grayscale then back to BGR for consistency

### Error Handling
- Graceful handling of None inputs with placeholder images
- Type coercion for silhouette (float32 [0,1] -> uint8 [0,255])
- Frame format auto-detection (grayscale, BGR, BGRA)

### Keyboard Interface
- cv2.waitKey(1) for non-blocking input
- m key cycles: 0 (Both) -> 1 (Raw) -> 2 (Normalized) -> 0
- q key returns False to signal application should exit


## Task 3 Decisions

### Type Annotation Choice
Used `object | None` for `_visualizer` attribute rather than importing `OpenCVVisualizer` type to avoid potential circular import issues and keep the module structure clean. Runtime type checking via `getattr` is used for the `close()` method.

### EMA FPS Parameters
Selected alpha=0.1 for EMA smoothing as it provides a good balance between:
- Responsiveness to FPS changes (not too sluggish)
- Noise reduction (smooths out frame-to-frame variations)

### Visualization Payload Structure
The payload dict structure was designed to match the `OpenCVVisualizer.update()` signature:
```python
{
    "mask_raw": UInt8[ndarray, "h w"] | None,
    "bbox": tuple[int, int, int, int] | None,
    "silhouette": Float[ndarray, "64 44"] | None,
    "track_id": int,
    "label": str | None,
    "confidence": float | None,
}
```

### Error Handling in run()
Frame processing errors are caught and logged (not raised) to ensure the visualizer loop continues even if individual frames fail. This maintains the real-time display even during transient errors.


## Task 5 Decisions: YOLO Model Path Relocation

### Decision: Model Path Structure
Decision: Move yolo11n-seg.pt to ckpt/yolo11n-seg.pt
Rationale: The ckpt/ directory already exists and contains ScoNet checkpoint

### Decision: Path Reference Style
Decision: Use relative path ckpt/yolo11n-seg.pt for CLI defaults
Rationale: CLI tools run from repo root; test uses absolute path via REPO_ROOT

### Decision: Preserve CLI Semantics
Decision: Keep existing CLI option names and only change default value
Rationale: No breaking changes to existing scripts or user workflows