Memory-budget scheduler: a sticky worst-case ratio double-scales sizes already raised to measured values

In a job scheduler that admits workers under a total RSS budget, each worker's need was computed as predicted_size x (worst measured/predicted ratio seen so far). To keep the run safe against a stale prediction, the ratio was made sticky. But a stopped and requeued worker already has its size raised to the measured value, so its need became measured x ratio (5-15x), and every requeued worker ran alone. On stand-in workers this took wall time from 28 to 46 s and from 15 to 29 s. The fix: keep the original prediction separately, apply the ratio only to that (need = max(orig_pred x ratio, current_size, floor)), and compute the ratio only over workers that have a real prediction, not a placeholder default. To catch this in testing, make a stand-in step up in memory after it looks stable, so it gets stopped, then grep the start log for how requeued workers were counted.