SegmentWindow

Qualified name: algan.animation\_timeline.timeline.SegmentWindow

class SegmentWindow(rows, endpoints, index0, index1, weights, cache_key)

Bases: object

One mob’s rows of one attribute over one frame window, described as endpoint states plus per-frame interpolation instead of a materialized [T, rows, C] buffer.

Produced by AnimationTimeline._describe_segment_windows() for attributes that opted in (AnimationTimeline.enable_segment_windows()) when the window’s writers are all plain recorded assignments – currently Surface.color_texture, whose per-frame materialization is a whole image. Frame t’s state is endpoints[index0[t]] + (endpoints[index1[t]] - endpoints[index0[t]]) * weights[t]; the weights are the exact per-frame rate-function values the dense replay would have computed (same tensors, same ops, same shapes), so the only difference from the dense buffer is the lerp’s re-association ((post - pre) * a against the replay’s change * a) – at most an ulp per texel, which is why the consumer’s flip is a qualified exception rather than byte-identical.

cache_key identifies the endpoint stack across batches (it is a tuple over the edit records the endpoints came from), so consumers can reuse per-stack work (u8 provenance, the stacked upload) while the same animation spans several frame windows. It is None when an endpoint was read from the mutable current state (the window runs past the last edit), which must not be cached.

Methods

evaluate

The window's dense [T, rows, C] state, for fallback fills.

Attributes

rows
endpoints
index0
index1
weights
cache_key

evaluate()

The window’s dense [T, rows, C] state, for fallback fills.

Only the escape hatch reads this (an updater discovering the mob mid-window, see AnimationTimeline.trace_updater_mob_access()); ordinary consumption samples the endpoints in the render kernel.