fix(middleware): skip raw tool-call fallback when invalid view carries the same call

[Baldwinzc]

Why

DanglingToolCallMiddleware._message_tool_calls collected the raw additional_kwargs["tool_calls"] payload whenever the structured tool_calls list was empty — even when invalid_tool_calls was non-empty and carried an entry with the same id. The raw payload is a fallback serialization of the same calls: the OpenAI serializer reaches for it only once both structured views are empty, which is exactly the gating _normalize_tool_call_ids documents and implements. Collecting it alongside the invalid entry counted the call twice, and _build_patched_messages then emitted two placeholder ToolMessages for one tool_call_id — the duplicate-id shape strict OpenAI-compatible providers reject with HTTP 400, which is the failure this middleware exists to prevent.

(Defensive consistency fix: the internal gating between _message_tool_calls and _normalize_tool_call_ids disagreed; standard langchain_openai never produces this shape, but patched OpenAI-compatible providers do write raw tool_calls into additional_kwargs, so the duplicate-placeholder path is reachable in principle.)

What changed

• The raw payload is now collected only when both structured views (tool_calls and invalid_tool_calls) are empty, matching the documented serializer behavior and the gating in _normalize_tool_call_ids. The raw-only fallback case is unchanged.

Surface area

• Frontend UI
• Backend API
• Agents / LangGraph — DanglingToolCallMiddleware tool-call collection
• Sandbox
• Skills
• Dependencies
• Default behavior change
• Docs / tests / CI only

Bug fix verification

• Test path: backend/tests/test_dangling_tool_call_middleware.py::TestBuildPatchedMessagesPatching::test_raw_fallback_is_skipped_when_invalid_view_carries_the_same_call
• Red on main, green on this branch: yes (fails on main with 2 placeholders for call_x; passes here with exactly 1)

Validation

cd backend
PYTHONPATH=. uv run pytest tests/test_dangling_tool_call_middleware.py  # 64/64 pass
uv run ruff check / ruff format --check                                  # clean

AI assistance

Tool(s) used: Claude Code

How you used it: AI located the inconsistency during a codebase audit, wrote the fix and test; I reviewed the change and verified the red/green test run locally.

• I've read and understand every line of this change and take responsibility for it — it's not unreviewed AI output.