BUG PLEXOR TEST: Routing stuck on Opus after escalation - no downgrade

Bug Bash - Plexor Routing Stickiness

Severity: CRITICAL 🔴

Description:
Once routing escalates from Haiku to Opus, it appears to get "stuck" and continues using Opus for ALL subsequent requests, even trivial continuations that Haiku handled fine before.

Evidence:
| Timestamp | Model | Cost | Response |
|-----------|-------|------|----------|
| 20:03:00 | Haiku | $0.028 | Full edit response |
| 20:03:03 | Opus | $0.140 | "continue fixing..." (9 tokens) |
| 20:03:07 | Opus | $0.140 | "continue fixing..." (9 tokens) |
| 20:04:56 | Opus | $0.143 | "add maxlength to website field" |
| 20:04:59 | Opus | $0.143 | "add maxlength to the website field" |

Total Opus waste: ~$0.57+ for responses averaging <10 words

Expected Behavior:
After short/simple completions, routing should downgrade back to cheaper models like Haiku.

Root Cause Hypothesis:

1. No downgrade logic after model escalation
2. Session-level model caching that doesn't reset
3. Missing output-quality-based routing feedback loop

Impact:

• Cost is 500% HIGHER than passthrough
• Completely defeats the 70% savings goal
• Users are paying premium prices for minimal output

---
Found during real-time Plexor forensic analysis session.

@claude and @codex please review this pr