fix: Add warning for embedding configuration mismatch

[amihos]

Problem

When configuring OpenMemory with an external embedding provider (Gemini, OpenAI, etc.), there's a critical configuration mismatch that causes semantic search to completely fail:

• Storage uses embedMultiSector which respects OM_EMBED_MODE → uses external provider (768d vectors)
• Query uses embedForSector which respects OM_TIER → defaults to "hybrid" (256d synthetic vectors)

This vector space mismatch results in all similarity scores being ~0, making semantic search non-functional.

Root Cause Analysis

The issue occurs because:

1. embedMultiSector (storage) checks env.embed_mode to decide embedding method
2. embedForSector (query) checks tier variable to decide embedding method
3. Default OM_TIER=hybrid causes queries to use gen_syn_emb() (synthetic)
4. But with OM_EMBEDDINGS=gemini, storage uses actual Gemini embeddings

These are incompatible vector spaces - cosine similarity between them produces essentially random/zero results.

Solution

This PR adds:

1. Startup warning when configuration mismatch is detected - warns users to set OM_TIER=deep
2. Model upgrade - Updates Gemini from deprecated embedding-001 to text-embedding-004

Example Warning Output

[CONFIG] ⚠️  WARNING: Embedding configuration mismatch detected!
         OM_EMBEDDINGS=gemini but OM_TIER=hybrid
         Storage will use gemini embeddings, but queries will use synthetic embeddings.
         This causes semantic search to fail. Set OM_TIER=deep to fix.

Testing

Verified fix by:

1. Setting OM_TIER=deep with OM_EMBEDDINGS=gemini
2. Storing test memory with unique keywords
3. Querying for those keywords - memory now correctly appears as top result with high similarity score

Alternative Approaches Considered

1. Make embedForSector respect OM_EMBED_MODE - More invasive change, might break intentional hybrid setups
2. Change default OM_TIER - Breaking change for existing users

The warning approach is safest and most backwards-compatible while clearly informing users of the issue.

🤖 Generated with Claude Code

[nullure]

Hey amihos, thanks for your interest in OpenMemory. I appreciate your pull request but it has conflicts with a file. Please resolve it

[nullure]

LGTM