Any Console or Computer that uses a FM Synthesis (eg Yamaha YM3812 found in the Adlib, SoundBlaster) actually produces frequencies that can't be captured with 48Khz. The SN76489 (Sega Master System, Coleco, PcJr/Tandy 1000) can generate frequencies that are well outside human hearing (eg 111Khz.) So yeah, audio is a bit more of a crapshoot to get perfect because the original audio generation wasn't perfect.
One issue with ultrasonic artifacts is it can mess with certain sound capture devices and create audible artifacts based on the interference pattern between the sample rate and actual ultrasonic harmonics. So it is a very good idea to provide a low pass filter on the circuitry. This also makes accurate emulation of sound a PITA because there is no such thing as a "digital" lowpass filter. It is necessary to take arbitrarily many samples across the sample period, then average them down so they scale nicely into 48khz. For instance, emulation of the waveform at 192000kHz and downmixing the samples down 4x using average would create a cleaner 48kHz signal with much less artifacts compared to doing a "read and hold" operation at 48000 samples without the benefit of a low-pass.
If the signal's ultrasonic content is wildly fluctuating between adjacent samples at 48000 Hertz, many of these abberations will produce audible artifacts below 16000Hz. For instance, suppose a strong tone is produced with a frequency of 52000Hz. Taking realtime instantaneous samples at a rate of 48000Hz would produce an interferance pattern with a frequency of 4000Hz. This would be entirely audible as a high-pitched shrill tone. In fact this phenomenon can be heard oftentimes when dialing in the AM band on a radio. Adjacent Channels are spaced 10kHz apart and these interfere with each other producing a very noticable 10Khz pilot tone.
Interference patterns can easily be avoided in real world applications by simply inserting a low pass RC filter to the analog signal, typically 20kHz for HiFi audio gear, prior to performing an ADC conversion. But analog signals are continuously variable, and in theory have unlimited bandwidth but this is in practice limited by stray capacitance, resistance, inductance, and interference from outside sources picked up by the conductors that carry the signal. Digital signals are limited by bandwidth, thus emulation, and even FPGA recreations of said signals have limitations. Given the tricky nature of some old analog hardware, an exact recreation of the CPU, GPU, and other components in FPGA, might still benefit by having real filter circuitry applied to an analog signal prior to being converted back to digital. If the traces are kept short and isolated, interference and noise could be minimized.
Otherwise, I'm not entirely sure if bandpass filters could be adequately applied from within the FPGA. My understanding is they are mostly digital devices, even if they offer limited support for analog signals on the outputs. I imagine the harmonic oscillations created by analog sound circuits present on very old arcade hardware, and perhaps in some consoles, would be difficult to emulate accurately even on FPGA without resorting to wave samples, as earlier builds of MAME used.