feat: self-contained PoC notebook for intelligent CC suggestion

[bhuvan-somisetty]

Adds a single Jupyter notebook (poc_demo.ipynb) that demonstrates the full CC suggestion pipeline end-to-end, covering all three goals from issue #2.

The notebook only needs numpy to run - no TensorFlow, MediaPipe, or ffmpeg required. The ML inference calls are replaced with realistic sample data so reviewers can execute every cell without model downloads or a GPU. The decision logic, label mapping, SRT/SLS formatting, and evaluation are real working code throughout.

The pipeline runs in four stages, each explained in a markdown section before the code:

1. Sound event detection - simulates YAMNet patch scoring, filters speech labels, drops events below the 0.35 confidence threshold, and merges adjacent same-label detections.
2. Visual reaction analysis - simulates optical flow motion scores and MediaPipe face-shift scores for the reaction window (300–1500 ms after each event).
3. CC decision engine - category-aware weighted fusion with a +0.12 boost for high-impact events. Weights are 85/15 for high-impact sounds (trust audio), 30/70 for ambient sounds (require visual confirmation).
4. Output and evaluation - generates SRT and SLS files and measures precision/recall/F1 against a ground-truth JSON using IoU-based matching.

The sample data is chosen to demonstrate all filtering cases: speech suppressed, low-confidence events dropped, adjacent detections merged, ambient music and a dog bark suppressed by the decision threshold, and India-specific labels (Fireworks → [firecrackers], Tabla → [tabla]) preserved correctly.

Fixes #2