Request consumption appears abnormally high — possible double/triple counting

### Describe the bug

I'm using the 1x multiplier model (GPT-5.4), but request consumption is behaving
as if a 3x multiplier model is selected. Each interaction burns through quota at
roughly 3x the expected rate, despite explicitly choosing the lower-cost model tier.

This started approximately 3 days ago with no changes on my end, same model
selection, same workflows, same shell environment.

### Affected version

GitHub Copilot CLI 1.0.22.

### Steps to reproduce the behavior

Set model to GPT-5.4
Run a typical planing - implementing workflow, says costs 1x req, but feels like 3x, %'s are increasing way too fast.

### Expected behavior

Expected: 1 request deducted per interaction (1x model)
Actual: ~3 requests deducted per interaction — consumption matches a 3x model

### Additional context

_No response_