Mocap as a pluggable Step: model / strategy / preprocessing

[mikub97]

What

For a non-coding researcher the pipeline starts at raw video, not at a
time-series CSV. Karnatak already runs mocap -> time series in production, but
those scripts are the ecosystem's two worst contract violators:
opt/step_motion_to_timeseries.py has no argparse at all and hardcodes
INPUT_DIR, OUTPUT_DIR and ./config.json.

Mocap should be a Step like any other, with three axes left open for extension:

• model — MediaPipe today (pinned 0.10.20; newer macOS-arm64 wheels are
  Tasks-only and drop mp.solutions). YOLO / OpenPose / others later.
• strategy — how many people and how they are found. Karnatak splits one
  video into two half-frames; that is one strategy, not the law. Also needed:
  one person per video, multi-person detect-and-track, per-person crops.
• preprocessing — smoothing, gap filling, coordinate normalisation, velocity
  derivation. Currently inlined; should be a declared chain.

Where

• DIMS_Dashboard_Karnatak/opt/step_motiontracking_twoperson.py (765 lines)
• DIMS_Dashboard_Karnatak/opt/step_motion_to_timeseries.py (229 lines)
• DIMS_Dashboard_Karnatak/opt/requirements.mocap.txt
• Contract: docs/contracts/step.md

Scope note

Ship current behaviour as the default mediapipe model + split-frame
strategy, so Karnatak works unchanged. The point of this issue is the seam,
not a second model.

Acceptance

• Karnatak's pipeline produces byte-identical time series through the new Step.
• Adding a second model or strategy requires no edit to existing step code.
• Preprocessing is declared in config, not inlined.