Test-Time Scaling in the Wild: Why Exploitation, Not Exploration, Is the Bottleneck