BYOK is broken in a fatal way!

BYOK is broken in a fatal way!
While I was trying out GitHub Copilot App with BYOK against an local endpoint which is supported by VSCode (so the endpoint should be working just fine), at the start of an session I received an error stating falling back to auto because BYOK was not available, which is bad in itself but okay.
Still I immediately pressed abort, switched back in the model selection from "auto" to my local LLM provider, then resent my message. Then I saw no error and everything seemed to be working fine now. (I wanted to try it out again because it is not uncommon for my local LLM endpoint to be sometimes not available for a few seconds because of live adjustments.)
But now after several new messages submitted with still my LLM provider selected, I suddenly receive a message with "You've used up all of your AI credits.", are you kidding me ?!?
I even checked the integrated token usage bar and it is STILL showing 1% used! While the GitHub web UI seems to correctly show that used up all my tokens for this month. Form the usage dashboards of my local LLM provider I can confirm that GitHub Copilot App has actually used its Cloud models because no usage was logged locally in the last hour. But the UI of the app is just plain lying, I cannot state it in a better way!

Reproduce this in GitHub Copilot App 1.1.17 on Windows using a VSCode-compatible local endpoint: make the endpoint unavailable, observe fallback to auto, then switch back to the local provider and send messages. Compare the selected provider, integrated token bar, GitHub web usage, and local provider logs. Done means BYOK failures do not silently consume cloud credits and the app reports usage accurately.