feat: migrate embeddings from Gunicorn service to Ollama / BYOK providers

[jordanpartridge]

Summary

Replace the custom Python/Gunicorn embedding service (knowledge-embeddings Docker container at :8001) with direct Ollama integration and support for bring-your-own-key (BYOK) providers like OpenAI.

Current State

• Knowledge uses EmbeddingServiceInterface backed by a custom EmbeddingService that calls a Gunicorn Python server running BAAI/bge-large-en-v1.5 at http://localhost:8001/embed
• This is the same model Ollama already serves locally (bge-large) on the same hardware
• The Gunicorn container is a single point of failure with no scaling, no fallback, and extra operational overhead

Proposed Changes

1. Add Ollama embedding provider — call Ollama's /api/embed endpoint directly (or via laravel/ai if feasible in Laravel Zero)
2. Add OpenAI-compatible provider — support text-embedding-3-large or any OpenAI-compatible API with an API key
3. Config-driven provider selection — ~/.knowledge/config.json or env vars to choose provider:

   {
     "embeddings": {
       "provider": "ollama",
       "model": "bge-large",
       "url": "http://100.68.122.24:11434"
     }
   }

4. Kill knowledge-embeddings container — remove Docker service, update docs
5. No vector re-indexing needed — same bge-large model produces identical 1024-dim vectors

Context

• lexi-agent PR feat: add Redis caching layer for sub-200ms queries #105 already migrated to Ollama via laravel/ai
• the-shit/vector v0.1.1 is now the Qdrant layer (PR feat: wire the-shit/vector as official Qdrant connector #145, merged)
• Ollama is running on Odin at 100.68.122.24:11434 with bge-large pulled

Acceptance Criteria

• Embedding provider is configurable (ollama / openai / none)
• Ollama provider works with existing bge-large model
• OpenAI provider works with API key
• knowledge-embeddings Docker container is removed
• All existing tests pass, no vector migration needed