Nice, thanks for working on this. I tested your fork, and it works for me as well.
This is a typical passing:
[DEBUG] Max correlation: 0.9223 (threshold: 0.70), DC offset: (0, 0)
S 9346 -44.558327 0 451 426
P 9169 OPN 6551418 -14.75 3721 0.53
As seen, out of 451 detected preambles, it was able to successfully decode 426. It's in-line with my experience of HackRF.
Detection threshold: the lower you set the correlation threshold, the more likely it recognises random noise as valid preamble. On very low settings, it will recognise practically anything as valid frame. On the other hand, I indeed see worse correlation with the rtl-sdr implementation. You can get it into shape by not tweaking the resampling's filter bandwidth (just use resamp_crcf_create_default).
EVM: yep, it's far from being nice, but for a different reason (it's bad with HackRF as well, I think the symbol synchronizer is used with an incorrect filter).
Amplifiers: The "RSSI:-4.41127" means you're using practically the full scale of the radio's input.