Measure warm-then-fan-out on your subscription

The experiment is two functions over one interface: fan_out releases N threads through a barrier so the calls really start together, and each backend's ask returns the cache fields from one call. The decision that matters most is where the lease sits in the prompt. A cache entry ends at a block boundary, so if the lease and the question share one text block, every question is a different block and nothing about the lease can be reused. Putting the lease in the system prompt (through --system-prompt-file) and only the question on stdin gives Claude Code's system-prompt breakpoint something to cache.