In terms of a good compromise for mainstream gaming performance, general desktop/business application performance, and price/availability, the K6 and K6-2 would obviously have made more sense for most people than either the P55C or MII . . . though in the context of the PR400 (or 433, if you could even find one), the P55C would be a used market CPU and K6 either much faster new (and more expensive than Cyrix) or used market as well, so that historical context isn't quite that straightforward.
Anyway, in terms of "gaming" performance in general, Quake and Quake II may be more extreme examples than most since it's very heavily optimized at low-level for the P5 architecture rather than aiming at broader (or older) x86 optimizations or using even less specific (and usually less optimized) compilers. Quake II was one of the last of its kind too in terms of having a software renderer option that heavily optimized at low-level.
While Cyrix optimized software was next to non-existent, it still may have been possible for less optimized games to fare better in the comparison than the likes of Quake.
There's also the issue of software renderers vs acceleration with more variables there too. You've got the possible optimization issues there too (custom renderer, API, etc), and potentially a wider gap in performance than software rendering as FPU performance alone becomes more important than other tasks needed only for software rendering if the GPU isn't the bottleneck. . . OTOH, software renderers using MMX may widen the gap too with the MII's less powerful MMX unit.
In your tests, it looks like the DirectX based MDK fared a bit better on Cyrix chips vs Pentium than Quake, and directX based games/drivers tend to be less heavily (and less specifically) optimized in general, so this may be part of the reason.
I'm not sure how Unreal might compare overall, and I haven't taken any actual framerate test figures yet, but in terms of general play speed "feel" my Asus P5A-B system running a P55C, MII, and K6, all at 2.5x100 seemed to run pretty close to equal both in software and Glide (Voodoo 3 3000 16MB PCI). though 3DMark99 and Final Reality (software and D3D) showed a major lead for the K6 over the MII and P55C over K6. The K6-2 also didn't seem to be very noticeably faster at the same clock/bus speeds, though it technically should be given the 3DNow! support for Unreal, and purely subjective "look and feel" testing is pretty vague, so I'll have to do more definitive tests later on.
Your 686 benchmark compilation isn't primarily a gaming benchmark though, so expanding it to the necessary sampling for a general overview of late 90s games would be way more work than necessary . . . on top of what you're already doing. Though, on that note, I'm still curious about how slightly older games would compare on these CPU families, specifically the last generation of ALU-only 3D games.
Additionally, I should note that my MII-366 2.9V that I managed to get to 3x100 2.9V (apparently stable) seems to be OK with a typical late-gen K6-2/Celeron capacity heatsink/fan, though it does appear quite temperature sensitive. The MII-366 is supposed to be rated for 70oC max, but at 300 MHz it seems to become unstable somewhere around 55oC. (at/below 50o it's been totally stable so far though)
I also decided to chance it and go all the way to 3.5V with the MII, but achieved nothing stable beyond 3x100, with 333 MHz crashing part way through loading windows and 350 MHz not going past the BIOS.
Note: I'm not using a thermal probe, but going by the motherboard's onboard CPU temp sensor.
The P55C at 3.2V doesn't seem to be nearly as sensitive, but my IBM 6x86L-200+ does seem pretty temp sensitive. At similar temps the MII-2.9V-300MHz (around or below 50C), the 6x86L seems stable at 3x66 MHz at 3.3V (need to test more for the actual stability threshold), but 83x2 also requires a moderate voltage boost for stability (3.0V seemed OK) and 2x95 seemed to be the most finicky to get stable in terms of temp/voltage. Even 95 at 1x required voltage boosts and 100 MHz wouldn't even post reliably at any voltage.
Feipoa, you also mentioned trouble getting your K6-2/550 stable or getting 600 MHz stable at all on a K6. If it's anything like my situation with the Cyrix chips, then temperature sensitivity may be a huge factor there rather than just voltage. Given the stories I've seen with K6-2/550 instability at stock settings (or even 2.4V), the 70C max temp rating is possibly a bit generous, and ample cooling may be the limiting factor for that CPU more than voltage. (unless maybe you were willing to go beyond AMD's 2.5V max rating on the chip and into Cyrix -let alone IDT- levels of voltage on a 250 nm part, and even then, with 2.8+ volts, you'd need adequate cooling to deal with the added wattage anyway)
On an MVP3/4 based board, 124x4.5 or (more likely) 112x5, or roughly similar settings on ALI boards, might also be more achievable (and better performing) than 6x100, but overclocking the chipset with >100 MHz FSB speeds would add another variable to the equation.