[Lucas1479]

Candidate processing previously interpolated between head and tail registration matrices, changing scale, rotation and translation over time. This restores the legacy pipeline's invariant: align the first frame once and apply that exact transform to the whole take. The tail estimate remains QA drift evidence and no longer changes the spatial mapping of intermediate frames. The actual applied matrix is recorded; endpoint locks remain a separate explicit blend.

This runs in generation-side Process & QA before adoption into asset editing and the runtime graph. UI and CLI labels describe whole-clip registration. Old registered results become stale and require reprocessing; already placed material with registration disabled keeps its recipe. Read-only checks never rewrite raw takes, published frames or configuration.

Based on main after #10 merged optional whole-clip black-border cropping. This PR contains only the fixed-registration behavior, recipe freshness, labels, documentation and its tests.

Validation:

• Focused processing regressions: 69 passed. The real ORB tests cover changing source scale/position, one exact matrix across every output frame, fixed margin, optional crop, drift QA failure, blocked adoption and unchanged raw/published assets.
• Recipe tests cover registered/unregistered results and legacy fields, including current/stale status and no writes during reads.
• Two local legacy videos, 60 frames each, passed actual decode, optional uniform crop, fixed ORB/RANSAC registration, CPU alpha processing, endpoint locks and QA. Original hashes stayed unchanged. Private media is not included.
• All nine Edge browser suites and JavaScript syntax checks passed. No real provider calls were used.
• Full Python regression: 321 passed, 2 skipped because this Windows host cannot create symlinks. Windows and Linux CI passed on the previous head. Rebasing onto the merged Add optional whole-clip protection-border crop #10 preserved the exact Git tree; CI is running again for the rebased commit.