Using cache=true did not prevent calls to the models

[bzorn]

I just added another case in packages.json that has caching turned on.
When you run

npm run promptpex:paper-speech-tag-cache

two times in a row without any changes to the repo, the second time you run it all the results from the models should be cached. But I observed:

1st run: (will fill in)

2nd run:

 ⇅ 108 ↑63.8kt ↓5.9kt 351.34t/s 2.38$ 3.3m 

I was assuming that running with cache=true 2 times would prevent calls to the model.

[pelikhan]

Found a bug where we were loosing parameters for test evals.

[bzorn]

Appears to be fixed now.