fix: Postgres reorganizer fails when pgvector embedding is returned as string

[Kwizii]

Bug Description

When MOS_ENABLE_REORGANIZE=true and GRAPH_DB_BACKEND=postgres, the reorganizer consumer crashes while preprocessing add messages:

pydantic_core.ValidationError: 1 validation error for GraphDBNode
metadata.embedding
  Input should be a valid list [type=list_type, input_value='[-0.047..., ...]', input_type=str]

PostgresGraphDB.get_node(..., include_embedding=True) assigns the pgvector column directly into metadata["embedding"]. With psycopg2, that value is often a JSON string rather than list[float], so GraphDBNode(**raw_node) fails and reorganize never runs.

How to Reproduce

1. Run MemOS Product API with Postgres graph backend and MOS_ENABLE_REORGANIZE=true.
2. Add a memory via /product/add.
3. Observe reorganizer logs in _run_message_consumer_loop / _convert_id_to_node.

Environment

• Python: 3.11
• Operating System: Linux (Docker)
• MemOS: main branch
• Graph backend: postgres + pgvector

Additional Context

• I am willing to implement this myself and have a patch ready on fork branch fix/postgres-embedding-normalize.

[Memtensor-AI]

🤖 AutoDev has picked up this issue and started working on it.

Task ID: 3b7006ec87efa6ed
Working branch: bugfix/autodev-2270-20260822151104895
Target branch: latest dev* branch
Workflow: opsp (analysis → coding → testing → PR)

I will post the PR link here once done. If I need more information, I will ask in the comments.

[Memtensor-AI]

✅ AutoDev task 3b7006ec87efa6ed completed.

Summary: Fixes #2270 by normalising the pgvector embedding column inside PostgresGraphDB._parse_row. A new module-level helper _normalize_embedding(value) -> list[float] | None accepts None, list, tuple, and both bracket- and paren-style Postgres text forms (e.g. "[0.1, 0.2, ...]"). psycopg2 returns the pgvector column as a str unless a type adapter is registered, and downstream code such as the tree-text reorganizer's _convert_id_to_node constructs GraphDBNode(**raw), whose metadata.embedding is typed list[float] | None; the previous direct assignment triggered pydantic_core.ValidationError and reorganize never ran.

Because _parse_row is the single choke-point for every read path (get_node, get_nodes, get_all_memory_items, get_by_metadata, get_all_by_scope, export_graph), routing embeddings through the helper fixes the bug for every caller with minimal blast radius. Unrecoverable input logs a warning and returns None so reads degrade gracefully rather than crash. No new runtime or optional dependencies; only stdlib json (already imported) plus a bracket/paren fallback.

Testing: added tests/graph_dbs/test_postgres_embedding_normalize.py with 13 unit + integration tests covering every helper branch and an end-to-end reproduction that feeds _parse_row output into GraphDBNode(**raw). Verification: ruff check and ruff format clean; pytest tests/graph_dbs/ → 44 passed, 3 pre-existing skips; two pre-existing failures in tests/memories/activation/test_kv.py reproduce on the untouched base branch (upstream transformers.DynamicCache API change) and are unrelated. Branch pushed; specs archived to memos-autodev-specs main.

Base branch: main
Branch: bugfix/autodev-2270-20260822151104895
Commit: b9c9eb5
PR: #2275
Assigned to: @CarltonXiang
Reviewers: @wustzdy