SummarizationMiddleware._Find_safe_cutoff_point Preserves Orphan ToolMessages Causing Provider 400

The method collects tool_call_ids from consecutive ToolMessages at the cutoff point, then walks backward looking for an AIMessage whose tool_calls intersect with that set. The intersection check passes if any id matches, even if other ids remain uncovered. In mixed cases where one ToolMessage's originating AIMessage was already summarized away but another AIMessage matches a different id, the function returns an index that still includes the orphan ToolMessage.

Run the provided Python example with SummarizationMiddleware configured with trigger=("messages",4), keep=("messages",3) on a history containing a mixed AI/Tool pair where one ToolMessage id is orphan. The middleware will produce a preserved slice starting at the AIMessage with the matched id, but also include the orphan ToolMessage, causing the next model call to fail with 400.

The fallback (returning run_end) may discard valid AI/Tool pairs that were immediately before the orphaned group, leading to loss of context and potentially degraded agent performance. This is a trade-off for avoiding provider errors; consider improving the summarization trigger to keep more messages or add a post-processing step that removes orphan ToolMessages without altering the cutoff index. Rollback is straightforward: revert to the original intersection logic, but note that the bug will reappear.