fix: make RAG vector retrieval reliable

[Summer-Of-Walden]

Summary

• Configure Reader Agent vector queries for the active builtin or remote embedding model.
• Persist per-book embedding provenance and reject model or endpoint mismatches before semantic search.
• Preserve retrieval diagnostics when hybrid search falls back to BM25.
• Decode Tauri SQL's persisted embedding byte-array TEXT form so chunk cosine fallback reads the original vectors.
• Fix desktop sqlite-vec startup: use the native busy timeout API and preserve an existing table's dimension instead of deleting a different-dimension index on restart.

Root cause

Reader queries previously lacked an embedding service. After wiring it, the code exposed incompatible endpoint contracts and a Tauri SQL embedding decode issue. Separately, VectorDB startup failed on a result-returning PRAGMA and could discard an existing non-384-dimensional sqlite-vec table.

Validation

• Core TypeScript typecheck passed.
• Focused core RAG and database tests passed (8 tests across 4 files).
• Rust VectorDB unit tests passed (2 tests).
• Desktop development build started successfully and initialized sqlite-vec.
• Manual Reader Agent regression checks completed.

Scope

CLI/MCP builtin embedding remains intentionally out of scope; CLI/MCP remote queries receive provenance validation.

[Summer-Of-Walden]

Superseded by #646, which keeps the RAG reliability changes and adds the mobile embedding implementation without overlapping #645.