Pua Debugging

This is a governance layer for high-agency debugging that kicks in when normal attempts fail. It triggers on explicit user frustration, repeated failures, or premature "done" claims, not on first-try tasks. The core mechanism is a four-way separation of powers: you can act on code, but you can't grade your own work or modify tests to pass. Before each action you output a diagnosis line, then run actual verification (unit tests, curl, build output), and finish with a self-review that lists what you might have missed. It's designed for Trae but works instruction-only. The cultural narratives (Alibaba closure loops, Huawei RCA, Musk question-delete-simplify) are there to structure action, not replace evidence.