0.152.0: Terminal studies run remotely by default

humanish 0.152.0 runs a live terminal study remotely by default, as a hosted computer-use study
already ran: its process runs in a runner sandbox on E2B or Modal and keeps going when this
machine sleeps. A remote run on E2B no longer loses its runner about 4 minutes in. The runner was
made from E2B's 473 MB base template, and a longer run's process was killed for memory without
writing a result. It now uses E2B's desktop template, and a 15-minute remote Claude run peaked at
471 MB with 6.8 GB still free. The committed synthetic-new-user persona no longer tells web
participants they rely on terminal output, so Taskly benchmark results from this release start a
new series.

Changed

• A live terminal study on e2b-terminal or modal-terminal now runs remotely by default, as a
  hosted computer-use study does: its own process runs in a runner sandbox on that provider, so it
  keeps going when this machine sleeps, and observe --run <id> reconnects to it. It has no live
  Observer, as a local terminal run has none; its Observer is rendered when it ends. The runner
  receives the study file, persona files and subject.product.upload. The runtime key Codex reads
  (CODEX_API_KEY, else OPENAI_API_KEY) enters the runner's environment on E2B as well as Modal,
  because the run places it in the shell sandbox; the participant gets it as before. --local or
  execution.remote: false keeps the run and its key on this machine.

• The committed synthetic-new-user persona no longer declares accessibility_needs: clear_terminal_output. A web participant read "You rely on clear terminal output", and in 2 of
  30 recorded benchmark runs its report discussed a terminal the study never showed it ("No
  terminal was involved, so there was no terminal output to interpret"). The TUI self-study now
  uses a new cli-newcomer persona, which keeps the need. Every study that uses
  synthetic-new-user gets a new persona digest and participant prompt digest, and Taskly
  benchmark results before and after this change are separate series. Personas written by
  humanish init never declared the need and are unchanged.

Fixes

• A remote run whose provider SDK (modal or @e2b/desktop) is not installed is now refused with
  the install advice before it names a run. It used to create a run id first and then say its
  sandboxes were not confirmed stopped, though nothing had been created (#1986).

• A remote run on E2B no longer loses its runner about 4 minutes in. The runner was made from
  E2B's stock base template, which has 473 MB of memory. The run's own process grew past 300 MB
  within two minutes, and a longer run's runner was killed without writing a result; runs under
  about 4 minutes finished. The runner is now made from E2B's stock desktop template (7.9 GB,
  8 vCPU) with no desktop started, so it costs about what one more participant desktop costs.