mapper/
├── index.html              # HTML shell: layout, modals, styles (Vite entry point)
├── src/                    # Modular ES6+ application source
│   ├── app.js              # Entry point: init, routing, domain switching, event wiring
│   ├── domain/             # Domain data loading and registry
│   │   ├── registry.js     # Domain index + bounding box lookups
│   │   ├── loader.js       # Domain bundle fetching + caching
│   │   ├── questions.js    # Question management + coordinate lookups
│   │   └── video-loader.js # Video catalog loading + window extraction
│   ├── learning/           # Adaptive quiz + knowledge estimation
│   │   ├── estimator.js    # RBF-based Bayesian knowledge estimator
│   │   ├── sampler.js      # Adaptive question selection (info-gain maximization)
│   │   ├── curriculum.js   # Domain progression + difficulty management
│   │   └── video-recommender.js  # Knowledge-gap-based video recommendations
│   ├── state/              # Application state + persistence
│   │   ├── store.js        # Central state management
│   │   └── persistence.js  # localStorage save/load (versioned schema)
│   ├── ui/                 # UI components
│   │   ├── controls.js     # Domain dropdown, zoom controls, header
│   │   ├── quiz.js         # Question display + answer handling
│   │   ├── modes.js        # Quiz modes (auto-advance, easy, hardest, don't-know)
│   │   ├── insights.js     # Knowledge insights panel (strengths/weaknesses)
│   │   ├── share.js        # Screenshot generation + social sharing
│   │   ├── video-modal.js  # YouTube player modal + video list
│   │   └── progress.js     # Loading modal (centered spinner + progress bar)
│   ├── utils/              # Shared utilities
│   │   ├── math.js         # RBF kernel, distance calculations
│   │   ├── accessibility.js # ARIA, keyboard navigation
│   │   └── feature-detection.js  # WebGL/canvas capability checks
│   └── viz/                # Canvas rendering
│       ├── renderer.js     # Main heatmap renderer (articles, videos, trajectories)
│       ├── minimap.js      # Corner minimap with viewport indicator
│       ├── particles.js    # Welcome screen particle physics system
│       └── transitions.js  # Domain switch animations
├── data/
│   ├── domains/            # 50 domain JSON bundles + index.json
│   │   ├── index.json      # Domain registry (ids, regions, bounding boxes)
│   │   ├── all.json        # Combined bundle (all articles)
│   │   └── {domain}.json   # Per-domain: articles[], questions[], labels
│   └── videos/
│       ├── catalog.json    # 5,044 videos with window coordinates [{x,y}]
│       └── .working/       # Pipeline working directory
│           ├── transcripts/    # 5,400+ Whisper transcripts (.txt)
│           └── embeddings/     # Per-video sliding-window embeddings (.npy)
├── scripts/                # Python data pipeline (see below)
├── embeddings/             # Generated .pkl files (gitignored, multi-GB)
├── tests/
│   ├── algorithm/          # Vitest unit tests (estimator, sampler, recommender)
│   └── visual/             # Playwright E2E tests (quiz flow, video recs, sharing)
├── notes/                  # Session logs and implementation plans
├── public/                 # Static assets (logos, favicons)
└── .credentials/           # API keys + cluster passwords (gitignored)

wikipedia.pkl (250K articles, gitignored)
    ↓ generate_embeddings_local_full.py
embeddings/wikipedia_embeddings.pkl (250K × 768)
    ↓ generate_domain_questions.py (Claude Opus 4.6)
data/domains/.working/*-questions-batch*.json (50 per domain)
    ↓ embed_questions_v2.py
embeddings/question_embeddings_2500.pkl (2500 × 768)

data/videos/.working/transcripts/*.txt (5,400+ Whisper transcripts)
    ↓ embed_transcripts.py (full-document embeddings)
embeddings/transcript_embeddings.pkl (N × 768, one per video)
    ↓ embed_video_windows.py (sliding-window embeddings)
data/videos/.working/embeddings/*.npy (per-video, [N_windows, 768])

    ↓ fit_umap_joint.py (articles + questions + transcripts TOGETHER)
embeddings/umap_reducer.pkl + article_coords.pkl + question_coords.pkl + transcript_coords.pkl
    ↓ flatten_coordinates.py --mu 0.85 (optimal transport)
embeddings/*_coords.pkl (density-balanced [0,1] coordinates)
    ↓ apply_flattened_coords.py
data/domains/{domain}.json (articles + questions with flattened coords)
data/domains/index.json (bounding boxes from question 5th-95th percentile)
    ↓ project_video_coords.py + export_video_catalog.py
data/videos/catalog.json (5,044 videos × window coordinates)

User answers question → estimator updates knowledge map
    → video-recommender computes difference map (estimated - ground truth)
    → scores each video's windows against knowledge gaps
    → ranks videos by expected learning gain
    → presents top recommendations in modal
    → user watches video → marks watched → map updates