Sandbox mount redaction clears caller-owned exceptions in place

Finding: In openai-agents-python, per an external report, when a sandbox call fails and the manifest has configured mount authority, redact_mount_error_data raises a generic RuntimeError but first clears the original exception graph in place. This includes exceptions reachable through __cause__, __context__, args and attributes, which may be caller-owned. Reported on 0.22.3, 0.23.1 and main at 125efa029b4b with Python 3.13 (https://github.com/openai/openai-agents-python/issues/5337). Reported symptom: in the reporter's script, the caller receives the redacted RuntimeError as intended, but a saved exception ends with an empty __dict__, a None __cause__ and a missing custom attribute. A separate exception that was only in the __context__ chain also ends with an empty __dict__ (same issue). Reported impact: in a Temporal workflow, a saved ActivityError from a failed sandbox acquisition was re-raised later. It had lost its message and cause, so reading .message raised AttributeError in workflow code. The reporter says this made Temporal fail and retry the workflow task on replay. The same mutation is said to apply to other methods decorated with redact_mount_error_data (start, stop, shutdown, aclose, exec, apply_manifest and others) and to _discard_exception_graph calls in agents.exceptions (same issue). Cause status: the cause is the reporter's reading of the code, not maintainer-confirmed. The reporter says they may be missing a threat-model reason for the clearing. Reported mitigation: save a detached copy of the exception and raise a new copy each time instead of re-raising the original object. This is the reporter's own workaround and was not verified by KFA. Proposed, unshipped ideas from the reporter: raise the safe error with 'from None' and drop SDK references; clear traceback frames only; or limit instance-state clearing to SDK-created exceptions, or document the behavior. Fix status: the issue was open with no fix or release identified in the supplied sources. Limitations: this is a single external report with a scripted reproduction; there is no KFA reproduction and no maintainer response in the supplied content. Comments coverage may be incomplete.