When parallel tool calls trigger loop detection warnings, the warning
messages (role="user") were inserted between tool result messages
(role="tool"). This breaks the Anthropic API when routed through
OpenAI-compatible proxies (e.g. LiteLLM): the proxy groups consecutive
tool messages into a single user message with tool_result blocks, but
an intervening user warning splits the group, causing orphaned
tool_results and HTTP 400 "tool_use ids without tool_result blocks".
Fix: accumulate warning messages during parallel result processing and
append them after all tool results, preserving the consecutive grouping.
Closes#642
