With_fallbacks(Exception_key=...) Writes The Exception Into The Caller'S Own Input Mapping

In langchain_core/runnables/fallbacks.py, each fallback invocation directly assigns `input[self.exception_key] = last_error` on the original input object instead of creating a copy. This mutates the caller's dict because Python dicts are passed by reference.

1. Create a RunnableLambda that always raises ValueError. 2. Attach a fallback using `with_fallbacks(..., exception_key='error')`. 3. Pass a fresh dict `{'a': 1}` to `invoke`, `ainvoke`, `batch`, `abatch`, `stream`, or `astream`. 4. After the call, inspect the original dict; it now contains the extra key `'error'` with the exception object.

The fix assumes `input` is a dict; if a non-dict input is passed while `exception_key` is set, the copy operation will raise a TypeError (same as the previous code, but without side effects). If upstream code later relies on the mutated original dict, this change would break that hidden contract, but such reliance is unlikely and considered a bug. To rollback, revert the six lines to the original in-place assignment. For extra safety, wrap the copy with `if isinstance(input, dict):` to gracefully handle non-dict inputs.