v0.2.6 Dotient Update

Added Fovea-1, a brand new model — our best model yet. Way more accurate search results and better understanding of your files. Big leap in quality from the previous models.

Added proper folder import — you can now drag a folder from Finder onto the app, or use the new "Import folder" button in the add dropdown. It scans the entire folder tree recursively, finds all supported files, shows you a count, and imports them all at once. Brand new — before this, you could only drop individual files.

Fixed raw camera image decoding — RAW formats (ARW, CR2, NEF, DNG, etc.) were coming out blurry or completely wrong because the Rust image library was decoding the tiny embedded JPEG preview instead of the full sensor data. Now it uses macOS's built-in sips tool for proper demosaicing, white balance, and gamma correction. Also generates a 1920px WebP preview that the detail view can actually display (browsers can't show raw files directly).

Fixed the graph view so it actually makes sense — the old cluster algorithm tried to tear the graph apart one connection at a time, which just didn't work for big collections. Replaced it with the Jaccard algorithm: it looks at how many shared neighbors two nodes have. If they share a lot of the same connections, they're probably in the same cluster. If they barely share any, they're a bridge between clusters. This finally gives clean, meaningful clusters instead of one giant blob.

Graph threshold was lying — the code logged one number for the edge cutoff but actually used a completely different one. Fixed so the displayed edges and the cluster detection actually agree with each other.

Thumbnails aren't pixelated anymore on Mac — macOS WebKit was treating a CSS flag as nearest-neighbor scaling, making every thumbnail look like a blocky mess. Swapped it for proper smooth scaling. Also bumped the internal thumbnail resolution so they're sharper even before the CSS fix.

Search won't leave you with a blank page anymore — if your search returns zero results, it now shows the 3 next-closest items with their similarity scores so you can click directly to them instead of staring at nothing.

Search results were showing the wrong scores — the merge algorithm was accidentally overwriting the actual similarity score with a meaningless rank-based number, so everything looked like it had the same ~3% similarity. Real scores are preserved now.

Images load immediately when scrolling — there was a race condition where newly visible cards in the grid wouldn't trigger their image load. Fixed so images appear the instant they enter the viewport.

Graph stays crisp when you zoom the UI — zooming the interface made the canvas go blurry because its internal resolution didn't scale with the zoom level. Now it renders at full resolution at any zoom setting.

Feedback appreciated!