Bug 1685938 added some code to remove "big" gaps between CPU measurements and their associated timestamps.
The solution was verified experimentally on a few profiles during development, showing that artificial CPU spikes had gone from 15-30x down to 1-3x only.

But there is no automated tests to verify this, and catch regressions.
This seems a bit difficult, given we're at the mercy of the test hardware, OS function calls, test environment at the time, etc. How to avoid false failures?

To be investigated...