Module geometry 

2026-09-26: Video token geometry: tokens per temporal group, and the check that temporal groups scale with clip duration.

A clip’s absolute token count depends on the server’s --video-fps, which the benchmark neither sends nor reads, so the geometry leg asserts a relation between three durations instead of absolute counts.

Owner: bench, video. Invariants: none beyond the types.

Structs

Ratio

2026-09-26: What the geometry leg concluded.

Functions

check_proportional

2026-09-26: Check that temporal groups scale with duration, from the totals of three clips at 1x, 2x and 4x, and infer the group count and template overhead.

tokens_per_group

2026-09-26: Merged tokens one temporal group of a w x h clip occupies: the still-image count, vision::geometry::expected_vision_tokens.