PGGB is the direct result of lack of critical thinking. The author claims the error of -600dB or better for his reconstruction algorithm.
Just had a closer look at PGGB and yes, this is completely insane, throwing quad-double (256bit) math at simple audio problems. A solution in search of a problem.
For the fun of it and as I already have my own time-domain convolver in double precision, I've used a brickwall filter kernel made with rePhase for a test run of a 2x upsampler. With that filter kernel (Albrecht-11 term, -6dB @ 21500Hz for 88.2kHz target sample rate, 2048 taps, -300dB down at fs/2 and above) -- and not using any dithering for that test -- I could achieve sample identity for a full-scale signal in the 24bit integer domain for any frequency up to 21100Hz, that is, the upsampling contains
all the original sample values except for an ultra narrow range at the top end. For music tracks, with much lower HF content, the match would be even better. As such, it can't get any better than this, me thinks.... and it's already overkill ;-)