Having trouble parsing this... could you show an example waveform of the CT output signal? What sort of pulse width/duty cycle does this issue occur at?
Okay, I'll explain it in a bit more detail. It's easier to see in the simulation, so I'll refer to these waveforms, but I'll also attach two measurements showing it in the prototype. It starts getting problematic at peak current levels that would demand an on-time of below 600 ns (which is a duty cycle of 600 ns / 5 μs = 12 % before duty cycle loss through the resonant inductor).
First, an example how it looks good and works well, input voltage at maximum and output power at maximum. Green (I(L6)) is CT primary current, blue (I(L14)) is CT secondary current, V(Iprim) is the voltage across the CT sense resistor and V(Iprim_F) is the same signal after a filter.

Note the CT currents at the marker position (passive-to-active transition, dashed line), it goes negative (normal with PSFB*) with some ringing. I think this is what causes the problems at lighter load, at the next picture:

The negative peaks are of course still there, together with the ringing the signal also goes positive though. The rectification on the CT secondary is the usual single diode. Through this rectification only the positive part of the ringing comes through, which then, after filtering, causes a bump in the measurement signal at the beginning of the active phase. I have to use leading edge blanking to ignore that part currently.
Here are comparable measurements in my prototype. Yellow / CH1 is equal to V(Iprim) in the simulation, the voltage across the CT sense resistor. Green / CH2 is after the filter, Blue / CH3 is the voltage between the switching nodes and red / CH4 is the transformer primary current.
Output at 8,38 V, 90 A:

Output at 1,68 V, 18 A. This still works stable, but when I go lower than this, the controller outputs 0 peak current set point but it can't go lower because of the leading edge blanking:

Is your objective to have an adjustable output voltage with a wide range?
Yes. CC and CV mode down to 2% of the nominal values.
Is your voltage feedback able to work with such low voltages (a typical TL431+opto circuit wouldn't work).
I measure output voltage and output current with a 2 channel ADC over SPI. I tested that circuit and it works well too. It's also not what is causing my problems at the moment, the digital controller's output is stable and all.
To have it working stable, I have a leading edge blanking of ca. 600 ns at the moment. When I reduce it down to below 350 ns or so, the CMC-Comparator starts triggering on the first "bump" in the primary current signal too, which leads to erroneous jumps in the duty cycle.
I don't really know how to fix this now, I only see the options of getting rid of the leading edge oscillations (but how?) or reducing the effective duty cycle by skipping pulses/lowering switching frequeny/using burst mode.
Am I missing some option?
I don't think PSFB is particularly unsuitable for wide line/load ranges (as compared to LLC which is pretty bad). For light load conditions, are you emulating diodes with the synchronous rectifier FETs, or are you operating in "forced ccm" mode?
I made the above measurements with forced CCM. I also tried turning off the synchronous rectifiers together with pulse skipping at low loads, but that doesn't improve the situation, it even makes it worse, because in DCM the output voltage is even higher at the minimum duty cycle.
* This is a point I don't understand fully yet. PSFB have a small, negative input current spike after the passive-to-active transition (see first simulation picture). This has to go through the CT, but, as it is in reverse direction than the normal sense signal, the impedance during that time is very high as the CT secondary diode blocks, so the impedance is mostly that of the CT primary inductance, which is quite high (2 μH in my case).
Is this correct?
Does this cause problems, maybe even my problem here (the ringing)?
Why do all the application notes from TI, Microchip, ST, etc. do it like this then?
Is there a "better" way?
Thanks for the input so far.
20V and 2000A is pretty high power. Nice!
Only 200 A, but yes, thanks

It sounds like this might be a job for pulse skipping mode. The basic idea behind the approach is to replace a continuous series of tiny-and-don’t-work-right pulses with a few sparse large-enough-to-work-right pulses (hence the name).
I already have pulse skipping and it works, but at the moment not enough. I can make it skip even more pulses, but it already gets into the audible range rather loudly and I have to increase the output capacitance probably, so I am looking for other options as well.
Thanks.