For my 486 distro[see snacklinux.org], I use syslinux 4.07 due to similar issues. I never had any luck with syslinux 6.x, I’d recommend a similar path. It always seems funny to me when I see similar projects, claiming it runs on 486 hardware but rarely do I see people actually doing that, and just fire up qemu instead. Running Linux in a vacuum isn’t realistic, especially when we’re talking old hardware and configuring IRQs manually.
It is running some AMI BIOS variant with a copyright date of 1992, I currently don't have the exact version string around to compare with the ROM dumps on retroweb. vbindiff says the "F" and "M" images are identical and the "H" only has a few 1-byte differences, mostly typos in ASCII strings.
I've written a small boot sector program once that tries out memory and CPU information gathering techniques, so I know the INT 15h, E820h, E801h are not implemented but INT 12h and INT 15h AH=88h return something sane. When I have more than 16M installed, the later reports the full 31M of HIMEM, but I'm not sure how the ISA memory hole factors into this.
From what I saw glancing at the scanning code yesterday, syslinux 6.x should fall back onto AH=88h if AX=E820/E801 doesn't work. It's interesting to know that this worked in older SYSLINUX, I'm curious to check out what changed.
Based on the POST strings of your motherboard versus mine (https://theretroweb.com/motherboards/s/fic-486-jal-rev-c), we both have an AMI BIOS so I might be able to run a similar test for my board. You're right that syslinux 6.x should fall back if E820 doesn't work but that hasn't been my experience on my motherboard hence the reversion to 4.x, I can reliably boot with 4-16MB of RAM.