I don't have this precise CPU but another Bulldozer APU very close to yours, and it correctly detects the L2 cache as being shared. The easy fix is to add parameter -t 2 or use a manual config. Avoid the no-cache threads on those CPU, it hurts the performances.

Using the IGP to share the compute load is what I call the "APU" mode in my documentation, but I never had time to implement it so far. 

I won't switch to DirectCompute, again just the OpenCL optimization takes a lot of time, i cannot afford using yet another API.

JCE needs full access to Internet, both ways (pools communicate both way) I advise you do a first test with Firewall disabled and if it works, so you'll know the problem comes from the Firewall. You may need to grant access to the jce_cn_cpu_miner64.exe and also to the shadow attrib.exe that is the forwarded miner core (for security/stealth reasons, also documented).
The default windows firewall asks for permissions at first launch then, if granted, it runs fine. Maybe you use a 3rd party firewall and/or antivirus?

Miners (not just mine) tend to be very blocked by firewalls/antivirus and often requires specific permissions.