Yeah – just seen that: #20. Thanks!

Thinking about it some more: this is probably only part of the story. It explains why BOINC gets “CPU time since checkpoint” wrong, but not why tasks’ elapsed CPU time resets to zero after being stopped and restarted.
---
Edit: Having tested it just now, I can confirm Mad_Max’s findings: forcibly updating the last-write time of docking_out.chk does cause BOINC to update all the related internal stats – including “Elapsed CPU time at last checkpoint”, which is saved for use after a restart to initialise the CPU time. And looking at the BOINC client code, it is clear that this only gets done when the last-checkpoint time reported by the wrapper changes.