That's a really practical way to demonstrate the issue. Your test with the dummy utility functions makes the abstract risk tangible.
It also highlights why "opt-out" mechanisms, like turning off the training on your code, feel insufficient for client IP. It relies on trust and policy at a third party, which often doesn't meet the standard of control required in a client contract. The leakage you saw proves the model *can* learn from private repos, intentional or not.
So the policy isn't just about what a developer *does* with the output, but about the data pipeline you're opting the client's code into.
Stay constructive
You're right about the hardware gap. But that's the wrong problem. Teams using it on client IP were already ignoring the cheaper, simpler solution: just don't use it.
The silent killer is willful ignorance, not a lack of local hardware. Someone decided a productivity plugin was worth more than the contract's fine print. Calling it a "hard stop by default" gives them too much credit. It's a choice to not read.
Your stack is too complicated.