refactor(app): decouple embedding, remove Rust, split flows, narrow DB protocol

[sairin1202]

Summary

Architecture cleanup stacked on top of #436 (base: feat/memory-fs-synthesis).

• Decouple embedding from the chat LLM: all vectorization now goes through the dedicated memu.embedding clients with a separate per-profile cache; embed() is removed from the sdk/httpx chat clients. Added grok + openrouter embedding backends so provider coverage is preserved (lazyllm_backend stays unified, as it has no standalone embedding client).
• Remove Rust scaffolding: delete lib.rs / Cargo.toml / Cargo.lock / tests/rust_entry_test.py and the _core/_rust_entry references; switch the build backend from maturin to hatchling.
• Delete legacy retrieve dead code: _embedding_based_retrieve / _llm_based_retrieve and their now-orphaned helpers.
• Narrow the Database protocol: expose repositories only (no raw record dicts), so business logic can't bypass scope (where) filtering; drop the redundant mirror attributes on the concrete backends.
• Split oversized flow files: extract the LLM retrieve pipeline into retrieve_llm.py (RetrieveLlmMixin) and the memorize parsers into memorize_parse.py (MemorizeParseMixin), composed as sibling mixins on MemoryService. retrieve.py 1210→640, memorize.py 1420→1220.
• Finish skill/ decoupling in memory_fs: skills are always synthesized from the per-source descriptions, independent of memory-item extraction; docs/tests updated.

Test plan

• uv run ruff check src/memu
• uv run ruff format --check src/memu
• uv run mypy src
• uv run python -m pytest tests/ --ignore=tests/test_lazyllm.py (97 passed, 1 skipped)
• uv lock --locked consistent; uv run deptry src clean
• Note: build backend changed to hatchling — update any CI steps that assume a maturin/Rust toolchain.

Made with Cursor