Temporal-context event models

The first GazeForge temporal model is a deliberately modest context MLP. It gives the event classifier access to neighbouring gaze samples while keeping the implementation within the core scientific Python dependency stack.

train_context_event_classifier() constructs a symmetric window around every sample, using kinematic and missingness features. Windows are built independently inside each participant/trial group, so context never crosses a trial boundary.

The window is specified in milliseconds and converted to samples using the recording rate. For example, a 50 ms radius at 60 Hz becomes three samples on either side of the centre sample.

ai_classify_events_context() returns per-class probabilities, confidence, model/version metadata, and the effective temporal radius. It applies the same sampling-rate compatibility guardrail as the non-temporal event model.

Temporal windows are constructed positionally, so duplicate or non-unique input DataFrame indices do not alter row alignment. The returned classification table preserves the original input index.

This model is a temporal baseline, not a performance claim. It must be compared with I-VT and the Random Forest baseline under participant-held-out and dataset-held-out validation before any claim that temporal context improves event detection. CNN and transformer models remain later candidates under identical frozen benchmark splits.