Seems like libwdpi.so is different between 382_10050 and 384_10007. Unsure if it's because of model-specific code in there or something, but switching libbwdpi version seems to have resolved it. For now I will split that library by model on the SDK 6.37 platform.
I'll have to see if the RT-AC88U and RT-AC3100 versions are different as well. if they are, it means that library contains model-specific code in it, and I will have to provide separate versions per model.
I suspect that might be the problem I initially had with 382_18991, where on my RT-AC88U it would randomly fail to initialize at boot. And the behaviour would change whether the RT-AC88U was hooked to my cablemodem, or within my LAN to another upstream router, what you've observed might explain that difference as well.
I'll have to see if the RT-AC88U and RT-AC3100 versions are different as well. if they are, it means that library contains model-specific code in it, and I will have to provide separate versions per model.
No idea if it's the same thing, but on my fork there were cases where qos would attempt to start before the WAN was fully up and silently fail. You can make a qos-start script (there are two cases with args init and rules on traditional qos) to put some data in the syslog to check the sequencing.
I suspect that might be the problem I initially had with 382_18991, where on my RT-AC88U it would randomly fail to initialize at boot. And the behaviour would change whether the RT-AC88U was hooked to my cablemodem, or within my LAN to another upstream router, what you've observed might explain that difference as well.