Remove confidence from perception output and experiments

[pskeshu]

Context

The paper shows VLM self-reported confidence is uncalibrated noise:

• Confidence when correct: 0.867
• Confidence when wrong: 0.857
• ECE: 0.524

The harness derives reliability from session history (stability, temporal consistency) instead.

Current state

confidence is kept as an optional field in PerceptionOutput with a default of 0.0 so existing experiments don't break. But several experiments still actively use it:

• duration_aware.py — confidence-gated transitions (_confidence_gate function). The entire strategy is built around confidence thresholds. Needs rethinking — gate on stability or temporal context instead?
• ensemble.py — averages confidence across 3 votes. Could use majority agreement count instead.
• changegate.py — falls back to previous stage when confidence is low. Same replacement: use stability.
• descriptive_multishot.py — includes confidence in the reconsideration prompt template.
• run.py — prompts still ask VLM to return "confidence": 0.0-1.0 in JSON. Harmless but misleading.

What to do

1. Remove confidence field from PerceptionOutput entirely
2. Update prompt JSON schemas in all experiments to drop the confidence field
3. Rewrite duration_aware.py to gate on harness-derived signals (stability, temporal analysis) instead of VLM confidence
4. Simplify ensemble.py to report agreement count rather than averaged confidence
5. Fix changegate.py similarly
6. Update run.py and benchmark/metrics.py to remove any confidence-related metrics