fix(mcp): keep ToolRuntime injection for sync-wrapped MCP tools - #5164

Fix looks right. Worth extending the Notes block above to record the new contract: functools.wraps copies __name__, __qualname__, __doc__, __annotations__ and __dict__, and sets __wrapped__. Two consequences worth stating there — (1) typing.get_type_hints(wrapper) follows __wrapped__, so string annotations from modules compiled with from __future__ import annotations still resolve against the original coroutine's module globals; that is exactly why this works for mcp/tools.py and skill_manage_tool.py, where runtime: Runtime is a string annotation, and (2) anything introspecting tool.func now sees the coroutine signature instead of (*args, **kwargs). Separately, the wraps on the inner run_coroutine has no effect on the fix (only the outer sync_wrapper becomes tool.func), so it could be dropped to keep the wrap surface minimal.

Done — expanded the Notes block and dropped the no-op wraps on the inner run_coroutine. Wording now states the functools.wraps contract (copies __name__/__qualname__/__doc__/__annotations__/__dict__, sets __wrapped__) and the reason it keeps get_type_hints resolving string annotations for PEP 563 callers (mcp/tools.py, skill_manage_tool.py). Also noted not to replace wraps with a bare wrapper. Pushed as part of d92f2b8.

Name typo: funopatched reads like a mangled func-patched. Since this becomes the canonical regression-test ID, test_func_patched_mcp_tool_keeps_toolnode_runtime_injection would be clearer.

Renamed to test_func_patched_mcp_tool_keeps_toolnode_runtime_injection.

This pins the MCP/session-pool path, but the same annotation erasure affected every tool patched by _ensure_sync_invocable_tool: skill_manage (its func wraps _skill_manage_impl), batch_status, cancel_batch, list_background_tasks, cancel_background_task, review_skill_package, update_agent and list_uploaded_files are all async-only with a runtime: Runtime parameter, so they ran with runtime=None in sync callers too. I verified at this head that _get_all_injected_args(skill_manage_tool).runtime == 'runtime', i.e. the fix does cover them, but nothing pins it — a one-line assertion (or a small loop over those impls) would stop a future wrapper refactor from silently regressing per-user resolution for the built-ins.

Thanks for the careful read. On the built-in list: I verified at this head that all the built-ins you named (skill_manage, batch_status, cancel_batch, list_background_tasks, cancel_background_task, list_uploaded_files, update_agent, review_skill_package) expose runtime as a `get_input_schema()` field, so `_get_all_injected_args` resolves it from the schema channel and they were NOT affected by this bug — even a bare annotation-less wrapper on `func` leaves them at `runtime="runtime"`. The affected channel is only the MCP adapter tools, whose `args_schema` is a JSON dict with no `runtime` field and which therefore depend purely on the wrapped coroutine annotations (covered by the MCP regression test). I added `test_sync_wrapped_builtin_tools_still_resolve_runtime` as a pin that `make_sync_tool_wrapper` keeps `runtime` resolved for the built-ins too, so a future refactor cannot regress per-user resolution — but its docstring says what it actually guards. If you would prefer I also add an assertion that a bare wrapper drops the MCP path, that is already the red side of the MCP test.