CroqTuner

CroqTuner harness — robust FSM-driven GPU kernel tuning loop with guard flags, idea dedup, and compaction-safe state. Wraps ai-tune-from-current-best and ai-tune-from-scratch with mechanical enforcement of the tuning protocol. Use when the user asks to "tune", "ai-tune", "croqtune", "tune from best", "tune from scratch", or "start tuning".

Sweep ALL scenario shapes and adapt the reference best kernel (f16 iter143, e4m3 iter068) to each shape with up to 30 iterations per shape. Non-stop until every shape across all scenarios has a tuned variant. Use when the user asks to "tune from best", "adapt kernels", "sweep tune", or "ai-tune-from-current-best".

Tune typical shapes from scratch (baseline .co kernels) for >=150 iterations each. Covers shapes far from the reference that need full optimization. Non-stop sweep until all typical shapes are tuned. Use when the user asks to "tune from scratch", "full tune", or "ai-tune-from-scratch".