fix: skip embedding models when auto-detecting Ollama provider

[MarceloAdan73]

What

Fixes the Ollama provider auto-detection so it no longer picks embedding-only models (e.g. nomic-embed-text), which fail at runtime with "does not support chat".

Why

On a machine where Ollama is installed with an embedding model as the first installed model, detectProvider() selected ollamaModels[0] blindly. The CLI then errored on every file with a confusing 400 does not support chat, retried 3 times and crashed with a Node assertion on Windows. This is the first-run experience for new users, so it had to be fixed before any public launch.

Changes

• Added pickOllamaChatModel() in providerDetector.ts (all 5 agents) that filters out embedding/image models (embed, nomic, minilm, mxbai, bge-, gte-, e5-, clip) and returns the first chat-capable model.
• When only embedding models are installed, detection now returns null, so the CLI prints the friendly "Set up one of these options" instructions instead of failing cryptically.
• Added unit + integration tests covering: skip-embedding selection, null when only embeddings, and first-non-embedding fallback.

Verification

• npm test — 123 tests pass (10 files)
• npm run typecheck — clean
• End-to-end against a real Ollama with nomic-embed-text:latest, qwen2.5-coder:1.5b, qwen2.5-coder:7b: now detects qwen2.5-coder:1.5b and generates docs successfully