[manuc66]

What

Adds local tag suggestions with human review. After a few photos are tagged, the app proposes tags for the others, with a score and the evidence behind it. Nothing is assigned automatically: a suggestion only fills the label panel when the user accepts it.

This is the safe half of the "AI sorter" idea: a decision model that suggests tags, while the existing Publisher remains the only thing that fans files out into folders.

How it works

• Local descriptor — a MobileNetV2 1.0 (224×224) ONNX model computes a 1280-value visual descriptor per photo. 14 MB, CPU only, no network. The model's pooled features are used instead of the ImageNet logits; the classifier weights are untouched. Source, licence and SHA-256 are documented in ImgTagFanOut/Assets/Models/README.md.
• Cache — descriptors are stored in item_embeddings keyed by content hash and model version, like face analyses. A second scan only reads SQLite.
• Three signals — file name/path, faces already named on the photo, and similarity with already-tagged photos (cosine kNN, only when the library has at least 3 distinct tags). Similarity is deliberately conservative: it needs a dominant share of the neighbourhood and a margin over the next tag.
• Review — tag_suggestions stores at most three scored proposals per image with their source and evidence. Accepting assigns the tag and records the feedback; rejecting remembers the tag so it is not proposed again.
• UI — a suggestion strip above the faces, a Suggest tags command with progress/cancel, and a Next suggestion command that walks the pending queue.

Notable implementation points

• Tagged photos are processed first, so the comparison gallery is complete before an untagged photo is scored, independently of file order.
• A missing model falls back to path and face signals; after repeated decode failures the model is disabled for the rest of the run.
• The gallery is capped at 2000 descriptors to keep the cosine pass bounded.
• The scan tolerates missing/unreadable files and saves its work every 50 images.

Verification

• dotnet restore --locked-mode, dotnet build --no-restore: clean, 0 warnings.
• dotnet test --no-build: 178 tests pass.
• Coverage: 83.7% lines / 64.8% branches (floors: 78 / 55).
• New tests include engine/scoring units, repository persistence, scan (cache invalidation, cancellation paths, order independence, missing model), the real ONNX model on the demo portraits, and headless UI tests that render the strip and accept a suggestion.

Notes / limits

• The confidence is a ranking score, not a calibrated probability.
• Visual suggestions need a cold start: at least 3 different tags in the library.
• A rejected tag is permanent for that image (no reset yet).
• No OCR yet; it can be added behind ITagSuggester later without touching the UI or the storage.