Add an Architecture Decision Record for the embedding model selection

[jamjamgobambam]

Document the tradeoffs between different embedding models (OpenAI text-embedding-3-small, a local sentence-transformers model, etc.) and explain why the current default was chosen.

Relevant files:

• docs/adr/002-embedding-model.md

Estimated effort: 3–5 hours

[IlyasFardaouix]

I noticed you mentioned embedding models, but the recent discussion seems to be focused on error handling. Can you clarify how the embedding model selection is related to the retry logic in error_handling.py?