Add outfit photo feedback, regeneration, and general layering

[yixin0829]

Users can now regenerate and rate modeled clothing previews and complete outfit images from the UI. Every version stays saved, ratings and comments attach to the exact displayed image, and Undo restores the previous version for one minute after regeneration completes, including after reload.

• Add a complete Outfits collection alongside Wardrobe and shared photo controls with optional regeneration direction, thumbs-up/down comment dialogs, pending/error states, and server-enforced Undo. Clicking the selected thumb clears its rating and comment on the backend without reopening the dialog; feedback events remain in the audit history, and duplicate clicks and stale responses are guarded. Use the selected compact title-row design: 15px borderless icons, a transient unboxed Undo link, and 44px touch targets on touch devices. Dialogs open with the cursor in the text field, and restoring a photo is announced to screen readers.
• Retain originals, replacements, rejected/invalid attempts and undone images in an immutable, content-addressed image archive (data/photo-history/) with prompt, reference context and feedback provenance. Explicit ratings and regeneration/Undo events are recorded separately.
• Keep library.json and outfits.json pointing at each target's active photo after a regeneration or Undo, so agents and the UI read one current image. The replaced copy in data/imported/ is removed once the history holds its exact bytes. Agent import and outfit delivery archive every attempt and use the same configured wardrobe directory as the app.
• Generalize AI-prefilled, editable layer suitability across the original five garment categories, removing shirt classification. Keep exactly two wearing-mode previews for layerable pieces and one otherwise, with experienced menswear styling and suitable inner pieces such as T-shirts or hoodies. Hide Ways to wear for accessories, bottoms and shoes. Place Create layer look beside the toggle and shorten its helper to Layer over a T-shirt or hoodie.
• Preserve supplementary body references and modeled-photo direction. Include the title-below-photo change and consistent gallery framing/transparent thumbnail trimming for equal-size trousers.

Validation: all 37 automated regression tests and the production build pass on the isolated PR head, covering selected-thumb reset, neutral feedback after reload, unchanged first-click/switching dialog flow, failure/race handling, and category/CTA behavior. Behavior was checked by hand in the running app against a local stand-in image provider with synthetic images: importing a layerable and a single-look garment through crop, cutout and modeled review (including a failed generation and retry), feedback with comments, regeneration with a direction, Undo after reload and after the one-minute expiry, outfit regeneration, and the 390 × 844 touch layout. Live AI photo quality and physical-device behavior were not checked. Fresh requests to the existing tailnet deployment verified the current frontend bundle and backend neutral-feedback handling without changing saved user feedback. Browser verification of the latest changes was blocked by the local browser runtime's HRESULT 0x80070003 initialization failure.

Personal wardrobe data, identity references, generated photos, feedback and environment settings are excluded from the PR.