…t interface

`_execute_tool_calls_concurrent` has gained two attribute / call-site
dependencies that the lightweight `_Stub` test fixture didn't carry:

1. `self._tool_guardrails.before_call(...)` is now consulted before each
   tool runs, and the result's `.allows_execution` gate decides whether
   to invoke the tool. The stub had no `_tool_guardrails` attribute, so
   both interrupt-fanout tests crashed with
   `AttributeError: '_Stub' object has no attribute '_tool_guardrails'`.

2. `self._append_guardrail_observation(name, args, result, failed=...)`
   is called for non-blocked results to fold guardrail observations into
   the tool output. The stub had no such method.

Additionally, `_invoke_tool` is now invoked with `messages=` and
`pre_tool_block_checked=` kwargs by the concurrent path, so the test
fixtures' `slow_tool` / `polling_tool` callables raised
`TypeError: ... got an unexpected keyword argument 'pre_tool_block_checked'`.

Fix all three drifts:

- `_tool_guardrails` is set to a permissive `MagicMock` whose
  `before_call`/`after_call` return objects with `allows_execution=True`
  (these tests aren't exercising guardrails).
- `_append_guardrail_observation` is stubbed as a no-op pass-through.
- `slow_tool` / `polling_tool` accept `**kwargs` so any future
  `_invoke_tool` plumbing is absorbed without yet another rewrite.

The interrupt-fanout invariants the tests *do* exercise (worker-thread
tid registration, `is_interrupted()` propagation across
ThreadPoolExecutor workers, post-clear cleanup) are unchanged.

No production code change. Fixes the two assertion failures observed
on `main` (run 25250051126):

`tests/run_agent/test_concurrent_interrupt.py::test_concurrent_interrupt_cancels_pending`
`tests/run_agent/test_concurrent_interrupt.py::test_running_concurrent_worker_sees_is_interrupted`