docs: align storage TDD with merged multimodal seams

[principle-lgtm]

Summary

• Updates the storage TDD after admin-merging PR feat(multimodal): cross-modal types + trait seams (closes #146, #148, #153, #156) #185 so it references the merged multimodal EmbeddingEngine / VectorStore seams instead of proposing parallel EmbeddingModel APIs.
• Clarifies production readiness: HTTP embedding providers and Qdrant dense search are usable today; Candle local inference, LanceDB, payload indexes, and hybrid sparse+dense retrieval remain implementation work.
• Tightens roadmap items around trait adapters, LanceDB implementation, Qdrant templates, and using AIVCS / TransitionLog for durable audit while keeping Valkey ephemeral.

Test plan

• git diff --check

Related

• Follow-up to docs: TDD — AI inference KV & Rust-native RAG vectors (research) #189
• Incorporates state after feat(multimodal): cross-modal types + trait seams (closes #146, #148, #153, #156) #185