GeezerVison

Summary

Adds LiteLLM as an opt in provider in both places CodeGraphContext calls a model: natural language graph queries (/api/ai_query) and the embeddings behind ENABLE_VECTOR_RESOLVE. Set LITELLM_MODEL=anthropic/claude-sonnet-4-6 (or bedrock/..., vertex_ai/..., ollama/...) and CGC_EMBEDDING_MODEL=litellm/<provider>/<model> to reach any of them, using each provider's own credentials. Nothing changes unless one is set.

Changes

• viz/server.py: _call_litellm() call_llm() uses it when LITELLM_MODEL is set, and ai_query() accepts it as a configured provider.
• tools/indexing/embeddings.py: _LiteLLMEmbedder. Vector size comes from LiteLLM's model map, or one probe call for unmapped models, resolved only when the Neo4j index needs it; results are reordered and length checked.
• cli/config_manager.py: LITELLM_MODEL and LITELLM_API_BASE keys; CGC_EMBEDDING_MODEL litellm/... now validates.
• pyproject.toml: optional extra litellm>=1.97.0,<1.106.0.

Tests

• 29 new unit tests against a stubbed litellm, through the real call_llm, /api/ai_query, _get_embedder, the backend probe, EmbeddingPipeline.run and config validation. Neighbouring suites unchanged; ruff check and ruff format --check clean.
• Live against Vertex AI, on litellm 1.97.0 and 1.105.3:

[chat] vertex_ai/gemini-2.5-flash -> HTTP 200
[embeddings] litellm/vertex_ai/text-embedding-005 -> probe (True, 'litellm'), index 768, 5 functions embedded