Shift repo to ontology-first capability reference model

Reframe the project from a vendor feature tracker toward an ontology-first AI capability reference.

Highlights:
- add ontology, taxonomy, migration, scope, and schema documentation
- introduce first-class capability, provider, product, implementation, and model-access records
- add ontology validation tooling
- generate a capability-first site view alongside the existing feature-first dashboard
- decompose the old open/self-hosted bundle into active open-model and runtime evidence sources
- move the legacy local-models bundle into archive-only storage
- update build logic so archived sources are excluded from the active site build
- refresh public copy and generated docs to reflect the new capability-reference direction

This preserves the current dashboard during migration while establishing the ontology-aligned data model and archive boundary for future work.

Author: snapsynapse

CAPABILITY_PRESSURE_TEST.md

@@ -0,0 +1,220 @@
+# Capability Pressure Test
+
+Status: Draft 1  
+Last updated: 2026-03-07
+
+This document checks whether the draft capability taxonomy can coexist with the current 70-feature dataset without forcing unnatural schema decisions too early.
+
+## Current Dataset Snapshot
+
+Current feature count by platform:
+
+- ChatGPT: 12
+- Claude: 9
+- Microsoft Copilot: 7
+- Gemini: 14
+- Grok: 8
+- Local/Open Models: 10
+- Perplexity: 10
+
+Current feature count by existing `Category`:
+
+- `other`: 17
+- `coding`: 8
+- `agents`: 6
+- `integrations`: 6
+- `research`: 6
+- `voice`: 5
+- `image-gen`: 5
+- `local-files`: 4
+- `browser`: 3
+- `search`: 3
+- `video-gen`: 3
+- `vision`: 2
+- `cloud-files`: 2
+
+The category distribution confirms that coexistence is feasible, but the existing category system is not a clean substitute for a capability taxonomy.
+
+## High-Confidence Capability Areas
+
+These parts of the draft taxonomy already fit the current feature set well.
+
+### Real-time voice
+
+Clear current implementations:
+
+- ChatGPT Advanced Voice Mode
+- Copilot Voice
+- Gemini Live
+- Grok Voice Mode
+- Perplexity Voice Mode
+
+This area strongly supports splitting listening and speaking into separate capabilities even if they are often implemented together.
+
+### Visual understanding
+
+Clear current implementations:
+
+- Claude Vision
+- Copilot Vision
+- Gemini Project Astra
+
+This is a stable user-facing concept and should remain capability-first.
+
+### Image and video generation
+
+Clear current implementations:
+
+- DALL-E Image Generation
+- Imagen
+- Aurora
+- Designer
+- Sora
+- Veo
+- Grok Imagine
+
+These are already intuitive user-facing capabilities and map well.
+
+### Research, search, and action-taking
+
+Clear current implementations:
+
+- ChatGPT Search
+- ChatGPT Deep Research
+- ChatGPT Agent Mode
+- Claude Cowork Mode
+- Gemini Deep Research
+- Perplexity Pro Search
+- Perplexity Agent Mode
+- Comet Browser
+- DeepSearch
+
+These need many-to-many mapping, but they are editorially strong.
+
+### Projects, files, and persistent work
+
+Clear current implementations:
+
+- ChatGPT Projects
+- Claude Projects
+- Perplexity Collections
+- Gemini in Workspace
+- Copilot in Office Apps
+- NotebookLM
+- Memory
+
+This is another strong area for capability-first navigation.
+
+## Areas That Need Careful Modeling
+
+These are not blockers, but they are the places where schema pressure is most likely.
+
+### 1. Model access is not the same as a user capability
+
+Features that fit awkwardly into a capability taxonomy:
+
+- GPT-4 Access
+- Gemini Advanced
+- Grok Chat
+- Model Selection
+- Llama 3.3
+- Llama 4
+- DeepSeek-V3 / DeepSeek-R1
+- Mistral Large / Mistral Nemo
+- Mistral Small 3
+- Qwen 2.5
+- Qwen 3
+
+These mostly describe:
+
+- access to stronger base models
+- access to more choice
+- quality or reasoning differences
+- local/open model availability
+
+Recommendation:
+
+- Treat these primarily as implementation or constraint records for now.
+- Do not rush to create a top-level capability like "use good models."
+- Revisit only if a stable user-facing question emerges, such as "Can I choose which model I use?"
+
+### 2. Reasoning depth may be a modifier, not a primary capability
+
+Potentially awkward features:
+
+- Claude Extended Thinking
+- Grok Think Mode
+
+These feel more like quality or depth modifiers on existing capabilities than standalone capabilities.
+
+Recommendation:
+
+- Model them as capability enhancers or constraints before modeling them as first-class capability pages.
+
+### 3. Workspace/build tools often span multiple capabilities
+
+Examples:
+
+- Claude Artifacts
+- Gemini Canvas
+- ChatGPT Canvas
+- AI Studio
+- Claude Code
+- Grok Studio
+
+These are best understood as workspaces or implementation shells that support:
+
+- document creation
+- code work
+- iterative editing
+- reusable context
+
+Recommendation:
+
+- Map them to multiple capabilities.
+- Avoid making "canvas" or "studio" a capability category.
+
+### 4. Browser-like products mix surface and behavior
+
+Examples:
+
+- Atlas Browser
+- Comet Browser
+- Copilot Vision
+
+These combine:
+
+- interface surface
+- search/research behavior
+- action-taking
+- visual context
+
+Recommendation:
+
+- Treat browser products as implementations.
+- Map them into capabilities like `search-the-web`, `take-actions-and-run-tools`, and `see-images-and-screens`.
+- Keep surface-specific information in constraints.
+
+## Draft Outcome
+
+The current taxonomy appears viable for coexistence.
+
+What the pressure test suggests:
+
+- The capability-first direction is compatible with the current dataset.
+- The old feature-first site can remain operational during migration.
+- The biggest unresolved question is how to represent model access and reasoning quality without turning marketing terms into top-level capabilities.
+
+## Explicit Risks To Call Out Later
+
+These should be surfaced again if they start affecting implementation choices:
+
+- If model-brand pages begin dominating the capability-first IA, the editorial model is drifting backward toward feature-first.
+- If reasoning modes become top-level categories too early, the taxonomy may become vendor-shaped.
+- If surfaces like browser, desktop, and API are treated as peer categories with user-intent capabilities, the IA will get muddy.
+
+## Recommendation
+
+Proceed with coexistence.
+
+The next concrete step should be adding a thin capability mapping layer while keeping the current feature records and current dashboard untouched.