[AdityasagarR123]

Summary

Adds a benchmark in experiments/ that compares 10 regression models for gaze estimation

What's included

• run_experiment.py: stratified 80/20 split and leave-one-point-out (LOPO) evaluation
• synthetic_data.py: 9-point calibration data generator with configurable noise and head shift
• sanity_check.py: diagnostics for data, learnability, units, leakage, fold integrity, and memorization
• README.md: setup, commands, arguments, metric definitions, interpretation
• results/: outputs and diagnostic reports from the synthetic benchmark

Verification

• Noiseless LOPO error: 0.00 px (evaluation pipeline is correct)
• Noiseless linear fit R² = 1.0

Status and limitations

• Results use synthetic data. Model rankings are provisional and should not be used to select a production model until real calibration recordings are benchmarked.
• Random Forest's low LOPO error is an artifact of the 3×3 grid (each axis is trained separately) and is excluded from the verdict.
• No changes to app/ or tests/ in this PR.

Open items (not in this PR)

• Real calibration dataset (gitignored; requested from maintainers)
• Production centering in app/services/gaze_tracker.py, lines 580–587 (pending decision)
• Broken tests in tests/test_gaze_tracker.py and tests/test_routes.py (pending decision)

Reviewer notes

Start with experiments/README.md, then run python experiments/run_experiment.py to reproduce the default results.