When a fresh container is provisioned. "per-trial" (the DEFAULT) gives each
eval trial a clean service so trial N can never observe trial N-1's writes —
the repeatability property that makes a side-effecting eval trustworthy.
"per-arm" shares one container across an arm's trials: faster, but only sound
when the task is idempotent or reads only.
When a fresh container is provisioned.
"per-trial"(the DEFAULT) gives each eval trial a clean service so trial N can never observe trial N-1's writes — the repeatability property that makes a side-effecting eval trustworthy."per-arm"shares one container across an arm's trials: faster, but only sound when the task is idempotent or reads only.