ORT CUDA EP lazily binds on first sess.run; this caused validator's first /predict to eat cold-bind cost (30-300s in TEE-VM) and scheduler reaped the instance before activation. Now Miner.__init__ runs a no-op inference so on_startup blocks until GPU is hot. Also drop concurrency:4 (default 1; our miner.py is not thread-safe) and remove timeout_seconds:900 (not a Chute() kwarg, silently dropped).
- chute_config: exclude broad pool instead of include:pro_6000, remove tee:true (lower compute_multiplier, wider scheduler pool)
- miner.py: dummy inference in Miner.__init__ to warm CUDA kernels before first /predict
