Summary

• suppress duplicate tool executions within a turn without blocking legitimate repeated tools after an intervening tool
• prefer verified final coach feedback over transient meta-review chatter when handing feedback to the next autonomous player turn
• route CLI coach-feedback extraction through the shared core extractor

Why

Autonomous runs were being degraded by two concrete issues:

1. repeated tool dedup in a turn was too aggressive and could block valid later tool reuse
2. coach feedback handoff could capture transient review chatter like "Let me check ..." instead of the actual final critique, causing the next player turn to repair almost blindly

Notes

• the OpenAI/OpenRouter tool roundtrip fix is already present in main; this PR adds the remaining stability fixes on top
• local cargo build/cargo test on Apple Silicon currently still hit the existing upstream zune-jpeg ARM blocker on main; this is independent of the changes in this PR