Fixing a Personal-Mode Latency SLA Breach

Your personal-mode try-on service currently routes every request through the diffusion cascade to maximize quality. Monitoring shows p50 latency of ~8 seconds, well past the 5-second interactive budget from Step 1, and session-abandonment rate on the try-on screen is rising.

1. Diagnose the root design mistake, referencing Step 3's architecture decision.
2. Propose a concrete fix using the cascade pattern from the scalability deep-dive, and describe exactly what happens on the request path for a typical user versus one who lingers.
3. After the fix ships, what would you expect to happen to average garment fidelity for the typical fast-path user, and is that an acceptable trade? Justify your answer against Step 1's assumptions.