Ok, with cgminer-4.9.2, my AU3 (https://www.amazon.com/gp/product/B016CWBYJK/) gets 12.5Gh/s over several days.  With cgminer-4.10.0, it gets 8.5Gh/s over several days.  The older USB miners all get the same on both versions.  Both versions have all generic USB drivers enabled.  I get the same results on both my 256M arm controller and an x86_64 desktop running Fedora 25.  I'm also getting a lot more accepted hashes with 4.9.2 - several times more, which the modest increase in total hashing power doesn't explain.  [^1]

Conclusion: There may be a (subtle and hard to debug) problem with 4.10.0.  :-(

I will try to submit this to Fedora as an official package.  Fedora will ship using the system jansson lib.  4.9.2 autoconfig is unable to find the system jansson lib, but 4.10.0 does find it.  

Another question:

When systemd kills cgminer (either version) on the ARM system, cgminer spits out a summary as it shuts down which goes into the system journal.  Nice.  When systemd kills cgminer on the x86_64 desktop, cgminer just dies.  No summary.  No "shutting down" message.  Why?

[1]: over a 3 hour period, my tiny rig got 9 accepted shares, with a square root of 3.  So the changes in accepted work units are within reasonable variance for such a small rig.