Smoke-test every harness x input combination before a paid cloud run

A benchmark VM died a minute after setup. A pre-flight check step crashed with a KeyError: two producers of the same group-config JSON used different key names for an equivalent field, and the pre-launch smoke had run each input through only one of two harnesses, so that pairing was never exercised. Before paying for a remote run, smoke the real cross-product (each harness x each input producer) on tiny inputs, and read optional config keys defensively. Also design the driver so a failed or empty pre-flight/parity step degrades (skip those cells, label them) instead of failing the whole machine. Otherwise one bug throws away expensive setup like a multi-GB input copy.