[ling0322]

The macOS app now writes each picture's JSON record in the docs/image-metadata.md format (environment / request / result, new ImageRecord), reads it and the old Record back, and shows it in a column of the preview window. libwaifu gains waifu_version_json and waifu_engine_device, and the model manager's catalog JSON a versioned_name, for the fields only the libraries know. The preview's FitWindow is removed: the window opens at 1000×700 and resizes freely.

Also in the app

• The big picture dims and shows an expand sign under the mouse, so it reads as clickable.
• "Reuse these settings" restores the device the record names (not while a job runs; the device list is one Session.devices, used by DeviceCard too).
• Speech and conversion keep the old Record; the format only covers pictures.

Notes

• device is spelled as the doc spells it (cuda-cpu-offload, so metal-disk-offload), and is the device the engine resolved: auto is never written.
• img2img writes inputs.start_from as the file it was read from; a pasted picture is copied to output/inputs/ first so there is a path to name.
• The preview reads the record before its first frame; reading it after made the column arrive late and the picture jump. Letterboxing in the preview is accepted: shaping the window to the picture took more code than it was worth.
• Not done: the PNG tEXt chunks, the library writing the record itself, CLI / webui / WPF.

Tests

./build/unittest, cargo test (waifu, waifu-modelmanager, waifu-cli), cargo build -p waifu-tools and swift test pass; cargo fmt --all is clean. New: ImageRecordTests (layout, null revision, read back) and WaifuKit tests for the version, the resolved device and versionedName. No model test suite covers the files touched.

🤖 Generated with Claude Code