For posterity sake, I was finally able to track down an answer to this issue. The documentation is not aligned with the actual behavior of the API.

From Michael Gurch at Google

Apologies for the confusion surrounding this issue. I was able to speak to a member of the product team on this. They informed me that the limit is 1,200 phrases total across all PhraseSets referenced in a single recognition request. All the phrase sets are merged into one prior to the validation essentially, and then processed.

I have requested that they update the documentation related to quotas to align with this constraint https://cloud.google.com/speech-to-text/v2/quotas#adaptation

I further requested they update the migration guide from v1 to v2 to call out this change in constraints that was introduced as it can break existing implementations like it did to mine. https://cloud.google.com/speech-to-text/v2/docs/migration