Module structural 

Level-2 (structural) caption generation.

Detects and describes temporal patterns within each sensor channel:

• Trends – monotonically increasing / decreasing / stable segments identified by fitting linear regression over overlapping windows.
• Anomalies – significant peaks, spikes, and valleys found by a prominence-threshold peak detector.

Algorithm overview

1. Denormalise (T, C) array.
2. Downsample time axis: 1440 → 36 points (factor 40, average pooling).
3. For every channel in every group:
   a. Fit linear regression over windows of size 6, 8, 12 data points
      (with 50% overlap).  Classify each window as increasing / decreasing /
      stable using slope and range thresholds.
   b. Run prominence-based peak / valley detector on the downsampled signal.
4. Sample up to `max_insights` per category, format with random templates.
5. Concatenate all group captions.

Functions

generate_structural_caption

Generate a level-2 structural caption.