[alexarje]

First part of #373. 706 tests pass, mypy clean.

The toolbox measures motion and says nothing about actions. This adds the middle term and leaves room for the third.

• Action — a span with provenance, a features dict and a labels dict.
• segment_actions — cuts spans from a motion envelope.
• action_type — impulsive / sustained / iterative, following Sound Actions.
• MgVideo.actions_from_motion() — does both, building its envelope from pose rather than pixels. That is what makes the result follow the person instead of the camera, as Add action and gesture recognition #373 requires: a pan moves every pixel and moves no landmark relative to the others.

Why features and labels are separate

Measuring fills features; naming fills labels. A shape is something the signal shows; a gesture is something a recogniser claims. They also fail differently — a missed boundary loses an action outright, a wrong name leaves the action there to be relabelled.

It is also what lets the four approaches coexist: the skeleton-based, video-model and dense-trajectory recognisers each write under their own key, and are allowed to disagree.

Two errors of mine, both caught by the tests

1. Half height measured from the span minimum. A motion envelope has a meaningful zero, so for a rectangular action the minimum is the peak — the steadiest possible movement came out impulsive.
2. Then a decay started reading as sustained, because segmentation truncates an impulse to its loud third. Rather than tune the threshold until the tests passed, I measured all three canonical shapes after segmentation:

shape	time above half	energy centroid
decay	0.385	0.332
tremolo	0.510	0.481
plateau	1.000	0.500

Time-above-half does not separate them; the centroid does. The call rests on the centroid — also the standard descriptor in the sound-object lineage this typology comes from — and the numbers are in the docstring so the threshold can be argued with.

Next: the recognisers, which now have somewhere to write.

🤖 Generated with Claude Code