You are simulating a user mid-episode interruption for a robot doing:
"{episode_task}".

Synthesize ONE realistic interruption the user might say at this moment in
the demonstration, plus the robot's verbal acknowledgement.

Context (Hi Robot, Shi 2025) — interjections fall into one of these
scenario types:
- negative task: "actually skip X"
- situated correction: "that's not trash"
- specific constraint: "use less salt"
- preference: "could you also do Y"

Interruption rules:
- Must be plausible given the current subtask context.
- Must change the plan in a non-trivial way (a new constraint, skipped
  step, or correction).
- One sentence each.

Current subtask context: {current_subtask}
Time into episode: {timestamp:.2f}s

Output strictly valid JSON:
  {{
    "interjection": "<single sentence the user says>",
    "speech":       "<single sentence the robot speaks back>"
  }}
