It's somehow good you did those proxy tests, i don't always use mine when i do my test, to be in real-life condition, and even when I do, i check for netcode correctness, not for the displayed numbers, which are cosmetical. And i never tried your mix of GPU + paused CPUs.

About it, I will not fix it completely, because there are tons of cases, all being corner cases:
* cpu paused during fee session
* or unpaused
* or paused then unpaused, possibly several times
* some but not all gpu paused, or all, or unpaused, or both, possibly several times...

Each case would require a recompute of fee level on the fly, a lot of code for corner cases. A rig is supposed to mine with all its configured devices.
So i fix the stuck session, but one who start JCE with cpu+gpu and then pause the cpu will still pay for the cpu mining service, even if unused. I rewrite the message "CPU Devfee" into "Devfee for CPU" to be more explicit. I will also update the doc to say to use --no-cpu to disable the cpu, and not pause it forever. It would also make JCE start faster and allocate less memory.

Next version will have:
* Yellow share counter fixed (cosmetical fix)
* stuck fee session skipped for paused-forever devices
* relaxed netcode security for Lermite
* faster JCE start sequence
* reworked a bit the OpenCL
* No watchdog yet (postponed)

edit:

0.32k online

I included a primitive, undocumented and mostly untested watchdog: --watchdog to quit if a non-paused GPU thread is stuck. And --watchdog N to quit if total GPU (ignoring CPU) hashrate goes below N h/s, except if a GPU is paused, or the pool disconnected.

@1rV1N : if you redo your test with version 0.32k, you should get a perfect one-for-one match between the yellow hashrate and your proxy. And no change pool side, since the fixes were cosmetical. Don't forget to add --no-cpu, or let your CPU really mine.