Summary

• Mid-loop buffer drain: Between tool iterations in ToolExecutionCompleted, buffered user messages are now injected into the conversation history before the next FireLlmCall(). User corrections like "stop" or "that's already done" are seen by the LLM on the next iteration instead of being silently queued until the turn completes.
• Duplicate tool call detection: Tracks hash(toolName:argsJson) per turn in HandleToolCallResponse. When the same tool+args appears 3+ times, a system nudge warns the model it is repeating completed work. Fires once per turn to avoid nudge spam.
• Both mechanisms are additive — existing hard cap (MaxToolCallsPerTurn=30) and budget nudge (75% threshold) behavior is unchanged.

Fixes #350

Evidence from production session

Session D0AC6CKBK5K/1773983518.003979 — bot scaffolded a GitHub repo (done by iteration 5-6), then continued for 15+ more iterations redoing its own work (git ls-files 4x, re-creating .gitignore, re-committing). 5 user correction messages were buffered and never seen by the LLM.

Test plan

• MaxToolIterationTests — all 4 existing tests pass
• Duplicate detection fires correctly in existing test (visible in test logs: turn_duplicate_tool_detected tool=web_search count=3)
• Build and deploy to dev daemon, verify turn_mid_loop_buffer_drain appears in daemon log when sending messages during tool loops
• Verify turn_duplicate_tool_detected fires for repeated identical tool calls in production
• Verify user "stop" messages are honored mid-loop