Harden saves: retry on commit conflict so answers are never dropped

Fix participant-file truncation: serialize + never-shrink saves, blocking final save

Select lowest-occurrence (filename,type) trial per image_id (first on ties)

Nest state.json as {image_id:{filename:{type:count}}}; keep one image_id per participant

Balance (filename,type) exposure per image_id using results-derived counts

Switch to access-code-based registration; save as <CODE>.csv

Update CSV with modified captions

Update app.py

Restore CSV (overwritten by mistake) with new test rows

Add participant registration, slugified per-participant CSVs, and resume logic

Normalize captions and add test trials to the study

Allow Gradio to serve images from IMAGE_DIR

Remove __pycache__

Remove accidentally uploaded venv

Initial deployment of caption preference study

initial commit