chore: bump all dependencies to latest versions

[LukeMainwaring]

Summary

Routine dependency refresh (via /updating-deps) bumping all backend and frontend dependencies to their latest versions, plus a refresh of the bundled AI library docs.

No source code changed — only dependency manifests, lockfiles, and docs. No deprecated/removed APIs are in use, so nothing required migration.

Backend (pyproject.toml)

Package	Old	New
pydantic-ai	1.95.0	1.106.0
transformers	5.8.1	5.10.2
fastapi	0.136.1	0.136.3
scikit-learn	1.8.0	1.9.0
sqlalchemy	2.0.49	2.0.50
uvicorn	0.46.0	0.49.0
pydantic-evals (dev)	1.95.0	1.106.0
ruff (dev)	0.15.12	0.15.16

Frontend (package.json)

27 packages bumped — notably ai 6.0.180→6.0.199, @ai-sdk/react 3.0.182→3.0.201, next 16.2.6→16.2.9, react/react-dom 19.2.6→19.2.7, @tanstack/react-query 5.100.10→5.101.0, plus radix-ui, lucide-react, motion, date-fns, axios, biome, ultracite, and type packages.

Docs refreshed

• docs/pydantic-ai-llms-full.txt (upstream restructured; URL now 301-redirects)
• docs/vercel-ai-sdk-ui.txt (28 pages)

Breaking changes

None affect this codebase. Grepped for every deprecated API:

• pydantic-ai's v2-prep deprecations (history_processors=, method-style result.usage(), tool_retries=/output_retries=, stream_responses(), MCPServer*) — the agent already uses the modern capabilities=[...] pattern and current VercelAIAdapter path.
• transformers 5.9–5.10 restructured Audio Language Models / removed Apex — this project uses ClapModel/ClapProcessor (audio embedding, not an ALM), unaffected.
• scikit-learn 1.9 deprecations target LogisticRegressionCV/SVC/TargetEncoder — preference.py uses plain LogisticRegression + StandardScaler, unaffected.

Validation

• ✅ Backend pre-commit (ruff, ruff-format, mypy)
• ✅ Frontend ultracite check (62 files)

Follow-ups (not done here)

• Runtime smoke-test CLAP loading after the transformers bump (tests mock it).
• The /updating-deps skill's Phase 4 curl -o for the pydantic-ai docs needs -L/-fsSL — the URL now 301-redirects and silently wrote a 0-byte file (worked around this run).

🤖 Generated with Claude Code