v0.5.2

• A page is read before it is scrolled. A read takes in the whole page, so scrolling one nobody has read only
  spends steps: a lookup for Mongolia on a long country list scrolled until recovery ran out and stopped stuck
  with the answer on the page. Once a page is read, scrolling it goes ahead, as a list that draws more on scroll
  needs.

• A reply the provider cut short is an outage, not a truncation. JSON that ends mid-value short of the
  output cap is asked for again at the same cap; a second one ends the run unavailable, naming how many
  tokens came back. Only a reply that used the cap gets more room and, if cut again, is reported as truncated.

• A browser that cannot be reached yet is an outage, not a failed task. A dropped or refused connection
  when the run first connects, as a cloud browser still starting can give, ends the run unavailable, so the
  eval runs it again; any other connection failure is still an error.

• An eval waits out a rate-limited answer key. Fetching a task's expected answer from a public API now
  retries a 429, 5xx or dropped connection with the same backoff as a run, where it used to end the whole eval.