[DTL]

"16K 100p 4:4:4 HDR PQ BT2020 12bit H.268 for PAL"

Some more longread about >FullHD frame sizes:

Yes - 4K is not totally invisible. 4K and more samples per frame may be used for higer-quality digital visual systems but in the following way:

1. Digital imaging systems have very important internal design param of samples per degree of view. The 'classic industry' SD/FHD/4K/8K/(16K) form family of system-60spd. They designed to the close to critical spatial samples density around 60 samples per degree. Below 60spd the visual resolution quickly degrade, above 60spd visual resolution very slowly increases. Only 8K start to move some higher to about system-80 or more.
With FHD at system-60 we already reach limit of 'normal' angle of view of about 30 degree horizontal. It is even a bit more to normal (about 16 degrees) so left some side space.
Higher family members of system-60 like 4K/8K/16K are completely unusable for general broadcasting and require special wide-angle production (and displaying).

But system-60 also require very high quality engineering of samples usage in both compression (creating low sample count buffer) and decompression (restoring image from sampled form) stages. As today it looks we still not have perfect 2D linear upsampler mathematics (the 1D+1D sinc is not perfect for 2D and some 2D-oriented resizers like Jinc/EWA-Lz and others are really not perfect also - the elementary dot is not perfectly round at all sizes). So nothing really mathematically perfect once for ages to suggest to ITU to be accepted as industry standard upsampler for moving pictures digital imaging at this planet. May be even rectangular sampling grid is not completely perfect for 2D imaging and need replacement to other type of 2D sampling grid.

But requirements for upsampler significantly relaxed if we not try to reach peak possible sharpness of digital imaging (limited by Nyquist and Gibbs) and use some more samples to describe same detailed image. And to restore display visual sharpness we simply increase design samples density per degree of view above 60. Lets to 120.

So frame sizes above FHD may be used for different systems:
1. 4K system-60 - wide angle, standard per degree quality system.
2. 4K system-120 - normal angle of view, increased quality per degree of view system.
Same with 8K and more - like 8K may be 'normal viewing angle' or 'full viewing angle' (as from initial FHD) system-240 with close to no special requirements for upsampler (simple bilinear can be used or direct samples to screen pixels mapping).

It also connected with settings of UserDefined2Resize with adjustable s-param: For system-60 targets it is benefitically to keep some non-linearity and use low s-value around 2 to make 'halo/over/under shoots' thinner but it still leave some residual non-linear distortions. But allow to use higher-compressed low initial sample count 1920x1080 frame and more simple MPEG codec for same average bitrate and filesize.

For higher-quality system-120 and more it may be recommended to use 'unlimited' kernel size (s > 3..5 and more) and output 'full' width 'halo/over/under shoots' and got less residual ringing and other possible distortions. The visual sharpness of 'oversampled' system-120 and more is enough high so no thicker over/under shoots will degrade it.

But the not very nice side of 4K and more systems is much higher requirement for MPEG codec like 4x and 16..64 more initial datarate to compress. So the usage of 4K system-120 may be as 'bruteforce' solution of better quality for poorly engineered 'initial sampling compression' frame.

It looks close to audio systems from 44.1 kHz for very perfectly and precisely engineered old systems for all parts - from ADC with very nice LPF, very complex processing and high oversampling DAC with good knowledge in sinc() and modern 192 kHz system running well on very simple LPF and poor DAC still very fine. Just because 192 kHz workflow is very cheap now.

So for good old MPEG codecs the good engineered digital visual workflows with low frame sizes may be used for 'classic' system-60. It require additional efforts in design good sample-compressor to low samples frame size and complementary decompressor for displaying (upscaler). The higher samples per frame count systems like system-120 and system-240 may use very simple bilinear resize or direct samples to pixel mapping with also good image but require much larger frame size to operate.

"There's no way it would be able to sustain days of workload without rebooting etc"

It is not really badly nature of Windows and DirectShow - I have experience on broadcast running some handmade DirectShow file player with controlled image overlay via sample-designed DS filter in between file decoder and renderer running for weeks/months or may be years without reboot. And got no complaints on its stability over many years running until it was replaced by other more featured solution. So the DirectShow core in Win7 is not very bad even for broadcast-graded stability. But other parts like ffdshow or AVS are subject to test. Though to run AVS as DirectShow intermeduate RAW processing component may be much easier in compare with ffdshow design but require once again some new DS filter design (or extract from ffdshow).