Handle embedding fallback dimension drift

[RaviTharuma]

Summary

• track active vector dimensions and avoid writing incompatible fallback embeddings into an existing index
• degrade semantic search to fulltext when runtime provider fallback changes embedding dimensions
• add an HTTP MCP regression that reproduces quota-triggered API fallback without breaking session search

Test Plan

• bun run test -- tests/embedding/provider.test.ts tests/memory/vector-stability.test.ts tests/search/orama-store-semantic.test.ts tests/integration/serve-http.test.ts tests/integration/high-risk-fixes.test.ts tests/integration/http-embedding-fallback.test.ts

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[AVIDS2]

I reviewed this one manually. The runtime fallback + dimension-drift handling is solving a real correctness issue in the current main branch, and the approach is coherent across provider init, observation writes/backfill, reindex, and search fallback. The regression coverage is also pointed at the right failure modes. Approving and happy to merge this as-is.