These three were dropped with the null-effect batch, but their trajectory metric collapsed for a fixable reason rather than a property of the case: the deleted object was excluded from both sides of the ratio, so failing to delete cost nothing and the denominator fell to tracker noise. The metric now charges a kept object its distance to the nearest frame edge on both sides, and scores a timed delete as a trajectory before the edit frame and a presence check after it, so these score normally. The other three stay out: their recorded transforms drop the object from frame 0 while the render removes it on schedule.
Drop 6 null-effect partway DELETE cases (132 -> 126 edits)
These six _after_impact cases delete an object after its interaction has finished, so the remaining objects trajectories are identical with and without the edit (measured divergence 0.00 px). The trajectory metric normalises by that divergence, so the ratio divided by tracker noise and produced meaningless values. Removed from the manifest, splits, thumbnails and scenes.
domino_chain: retime the timed DELETE to frame 24 (mid-fall)
edit_remove_domino_1_after_cascade (DELETE domino_1 AT FRAME 48) is replaced by edit_remove_domino_1_mid_fall (AT FRAME 24). At frame 24 the first tile is still falling at 38 deg, three frames after it first touches tile 2, so the cascade still completes but lags the source by 6-8 frames. 132 edits / 152 clips.
