When an exception occurs during f(*args, **kwargs), the span is exited via the finally block in _wrap_message_create with span.__exit__(None, None, None). This discards the exception context that should be passed to __exit__(exc_type, exc_val, exc_tb). While _capture_exception captures the error to Sentry, the span itself won't have the exception properly associated with its exit, potentially losing tracing context for error correlation.

In _sentry_patched_create_sync, the span is entered with span.__enter__() at line 626, but span.__exit__() is only called at lines 680 or 683 inside a with capture_internal_exceptions(): block. If an exception occurs in _get_token_usage(), content block iteration, or _set_output_data() before reaching span.__exit__(), the exception is silently swallowed by capture_internal_exceptions() (which returns True to suppress exceptions), but the span is never closed. The wrapper's finally block only closes spans with INTERNAL_ERROR status, which is not set for internal SDK exceptions. This causes span resource leaks.