H3 Attention Measure

ALPHA, hypothesis under test. Adds strength * log(1/duplication) to the pre-softmax logit of every TARGET VIDEO key, from the same hold map H3 Time Smear used to retime the clip. Held frames become several near-identical tokens, and plain attention gives that region several times the softmax mass; this is the correction for that, and only that. Audio, text, conditioning and reference rows are untouched. mode off = installs nothing. strength is a ladder: 1.0 is full compensation, 0.5 half, 0.0 none. A uniform hold map (all 1) is a no-op by construction, which is what the unit test gates on. Dense only: it refuses to install on top of a sparse-attention override, because a sparse pass drops keys and the measure correction would then be measuring the sparsifier.