The real-model acceptance test now builds a fresh fixture world per attempt
and runs the model-dependent turns over it. A reply the interpreter or agent
turns into a decline is a content failure and is retried (3 attempts); an
error, pending run, interruption or unapplied command is a fault and fails the
gate at once. Each failed attempt prints its assertion and a bounded excerpt of
the reply; the persona prose log now carries every failed attempt, the
acceptance-pass-rate line, and the passing persona prose. The recipe adds
--nocapture so a passing gate still shows its failed attempts in the step log.
Not verified yet: this commit has not been built or tested.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
6aec390f64