Hidden test counts and enforced model lists

What changed

• Run Tests now also grades the hidden cases and shows how many passed, so a candidate finds out before submitting. Candidates still never see the hidden inputs or expected outputs.
• Claude Code and Codex questions only run models from the recruiter's allowed list. Publishing stops and names any terminal question whose list has no model that can run it.
• Attempts that run out of time are now submitted and graded automatically. Webcam proctoring now starts, and in candidate review a skipped question no longer shows another question's code, tests, or score.