Segmentation Annotation Studio — annotation tooling, AI-assisted selection, exports, hardening & docs

[davramov]

Overview

This PR turns the tool into a full local segmentation-annotation studio for scientific imagery. It adds a complete drawing/editing toolset, AI-assisted selection (in-browser SAM), display enhancement, a Browse/Connect data workflow over Tiled and local folders, versioned saves, COCO and DINOv3/Lightly exports, install/startup + security hardening, a Lighthouse-driven performance/accessibility pass, and MkDocs documentation. Rebranded from "SAM3 Annotation Studio" to Segmentation Annotation Studio (it's now model-agnostic).

Annotation tools & editing

• Tools: polygon, magnetic lasso (livewire), rectangle, ellipse, brush, fill (paint-bucket), eraser, select, pan.
• Boolean-geometry editing (via polygon-clipping): clip-to-other-classes, merge-overlapping-same-class, and erase all preserve existing vertices at full resolution — regions tile flush with no gaps and don't erode on repeated edits.
• Eraser works on any shape kind, is radius-aware (carves on edge contact), rebuilds polygon vertices to match the cut, and can split shapes / open holes.
• Brush builds one shape from connected strokes; disconnected dabs become separate, individually-selectable shapes.
  Polygon/lasso node-undo: Cmd/Ctrl+Z removes the last node while drafting and reopens a just-closed shape to edit mode.
• Vertex add/delete (double-click), select-all with class/all scope (Cmd/Ctrl+A), region ops (merge/grow/shrink/remove-islands), copy/paste, brush cursor stays visible while space-panning.

AI-assisted selection

• Magic wand with two engines: Smart (SAM / SlimSAM) running in-browser via transformers.js (WebGPU→WASM fallback, offline-vendored weights), and a classic intensity wand.
• SAM prompts (click/box, shift-add, alt-exclude), other-class regions as auto "not" hints, and a "connected regions only" option to drop speckle.

Display, classes & workflow

• Display (view-only, export-safe): brightness/contrast, levels histogram, colormaps, gamma, CLAHE (adaptive contrast), sharpen; sliders debounced with an optional logarithmic scale.
• Classes: add/edit/recolor/hide, hotkeys 1–9, colorblind-safe palette, duplicate a class + all its shapes, quick-add from the annotation guide.
• Versioned saves + crash-recovery autosave; Insights QA dashboard; a Reference guide page for expert annotation instructions.
• Feedback button → Google Form with auto-prefilled app context; Docs & Feedback open as in-app iframes.

Browse, Connect & data

• Connect to a Tiled server or local folder; drag-and-drop ingest; typing a comma-separated class list pre-populates a dataset's classes.
• Browse with metadata facets, annotation-status, and star-rating filters.
• Exports: COCO (RLE + images + semantic/per-class mask PNGs, for SAM3) and DINOv3 / Lightly (images/ + masks/ label PNGs + classes.json); optional push-masks-to-Tiled.

Install/startup & security hardening

• start_all.sh: auto free-port finding for all services, uv auto-install, offline SAM vendoring; Tiled API keys auto-generated into gitignored .env and never exposed to the browser (all catalog/array access proxied server-side).
• New PROD=1 ./start_all.sh mode builds the optimized SPA and serves it from FastAPI (single origin, gzip).

Performance, accessibility & SEO

• Route-level code-splitting: the /connect entry dropped from a single ~980 kB chunk to a ~20 kB shell, with konva/AnnotatePage/polygon-clipping deferred; vendor manualChunks for caching; backend GZipMiddleware.
• Downscaled the logo (146 KiB → 11 KiB) + explicit dimensions (fixes CLS); WCAG-AA contrast fixes; form-control labels; role="option" on class rows; meta description + robots.txt.

Docs & testing

• MkDocs site (install, quick-start, per-tab guides, keyboard shortcuts, troubleshooting, and an expanded architecture page with design decisions and per-tool internals).
• Frontend unit tests (Vitest) and backend tests (pytest) covering geometry, clip/merge, mask↔polygon round-trips, CLAHE, export writers, and class duplication.

Not in this PR (planned next)

• The ipred AI/ML integration (feature compositions, CatBoost training, conformal prediction, encoder embeddings, prediction→editable-shapes) is planned and scoped but not included here.