fix(gallery): traverse image subfolders when finding orphaned db entries

[lstein]

Summary

Fixes data loss in scripts/gallery_maintenance.py (logic in invokeai/backend/util/gallery_maintenance.py) when an image_subfolder_strategy is configured.

When invokeai.yaml sets an image_subfolder_strategy (date, type, or hash), new images are written into subfolders of outputs/images (e.g. outputs/images/2026/03/17/<name>.png). The Clean Orphaned Database Entries operation only checked the top-level outputs/images directory for each image, so every image living in a subfolder appeared to be missing — and its perfectly valid database row was deleted. Users reported thousands of entries being wrongly removed.

Fix

• Add PhysicalFileMapper.get_all_image_filenames_recursive(), which globs the entire outputs/images tree (**/*.png) and returns the set of all image basenames found anywhere.
• Use that index in clean_orphaned_db_entries: a DB row is now only considered orphaned if its file is absent from the whole tree, not just the top level.

Why this is safe:

• Image names are globally unique UUIDs, so a set keyed by basename is collision-free across subfolders.
• Thumbnails are .webp and are excluded by the .png glob.
• The images-archive directory is a sibling of images/, so it is not traversed.
• Building the index once and doing O(1) lookups is also faster than the previous per-row os.path.exists stat calls.

Scope is intentionally limited to the orphaned-DB-entry operation, which is the one that destructively deletes database rows. The archive/thumbnail operations only move or create files and were left unchanged to avoid surprising side effects.

Testing

• python3 -m py_compile passes.
• Verified the recursive glob against a representative tree: flat, date-subfolder, and hash-subfolder images are all found; .webp thumbnails and the sibling images-archive directory are correctly excluded.

🤖 Generated with Claude Code