Something similar was done years ago using Socket 3 and a POD on an ASUS board.

Some So5 boards had *very* low FSB options (as they were basically using 486 PLLs). One I had myself: Spring Circle P5C01. It had options down to 16MHz. That might not sound slower than the other options but...

This is a board with SiS 5511 chipset and a 6205 VGA chip using UMA. So the memory bandwidth with 16MHz FSB is shared between CPU and VGA. It was *SLOOOOOW*, running Windows XP not quite as slowly as on my Pentium Overdrive I was running at 16MHz (16MHz 32b bus x 1 by removing fan), but being in same ballpark.

Ugh, sounds like torture to me. I recently used my TXP4 to test a K6 300 I have to make sure it works. And it does, but the difference between 6x83 with 256K L2 cache and 4.5x66 with only the motherboard's 512K L2 was night and day.

PLL modding when? 😜

Shuttle HOT-541 (430 FX) and HOT-553 (430 HX) have very flexible PLL. You may go down to 16MHz FSB.
Below is table (frequencies are measured with osciloscope). At 16MHz FSB motherboard was still able to boot-up with PCI Trio64V+
This is my HOT-553 with Pentium 120 (CPU multiplier set up as 1.5), I also have Cyrix 6x86 so I may try with 1x multiplier as well (in the future).

I just had a thought... while Cyrix 6x86 and MII class might have a functional 1x multiplier (Implemented multis can vary with time, fab, and OEM version, apparently identical ST/IBM/Cyrix versions of same model have shown differences.) ... it might not be slower than a P54 or P55 at 1.5x ... because that core is particularly fast on 16 bit code... anything you need about "486" fast is probably mostly 32bit by then... anything you want TurboXT/286 fast is probably 16 bit code. It might get brought more into line with L1/L2 kneecapping, as then the 50% loss of memory speed is gonna show.