Control motion in generated video

Fix the motion in this shot.

Diagnose first:
1. How many things move — subject, camera, background elements, clothing, hair. *Every additional moving element multiplies the failure rate.* Count them.
2. Whether the camera and the subject both move. If so, that alone explains most drift.
3. Clip length. Degradation compounds; the back half of a long clip is where things fall apart.
4. Contact and occlusion — hands holding objects, people passing behind each other, feet on ground. These are the specific hard cases.

Then rewrite:
- **One subject in motion.** Everything else explicitly held still.
- Camera static, or one simple movement, never both with subject motion.
- The action described as a single continuous movement, not a sequence. A sequence belongs in separate shots.
- Physical contact stated explicitly if it must happen, or removed from the shot if it does not.
- Duration cut to the shortest that serves the shot.

Then:
- Whether this shot should be split into two generations.
- What to accept as good enough. *Some of this is a model limitation, not a prompt problem* — say which, so I stop rerolling something that will not improve.
- Whether a still image plus motion in editing would be more reliable than generating the motion.