- New prompt_search.py (no ComfyUI imports): reads only the PNG 'prompt' chunk (tEXt/zTXt/iTXt), skips the large 'workflow' chunk, and extracts only text that feeds text-encoder nodes (following links to string/primitive/concat nodes). Model names, sampler names, filenames and other widgets are no longer matched, and filenames are no longer searched.
- Per-file results are cached by (mtime, size); repeat searches only touch new/changed files.
- gallery/list runs in a worker thread so large searches don't block ComfyUI's server; the scan stops when the client disconnects.
- Frontend aborts superseded search requests and ignores stale results.
- UI strings/locales and README updated; unit tests in tests/.