The issue was twofold, on one hand, using TNS on top of PNS meant the noise that got inserted was shaped by TNS, which is nonsense since the decoder generated the noise, not the encoder. This made PNS explode. The second, biggest issue was that using PNS in combination with any stereo tools resulted in noise leaking in both channels equally, ruining stereo imaging. So the best and only thing to do was to enable PNS only if the band in both channels is noise (or is sufficiently non-tonal and masked).