As long as the amount of float in the top pulley is a small fraction of the width between cogs, IMO, Shimano's approach is superior because it allows the chain to stay centered better and more quietly on a gear relatively well until a deliberate shift initiations derailment of the chain to another cog. As long as the float is a small fraction (e.g. less than 1/4th the gap between cogs) then an indexed shift will happen and complete just as fast. Yes, psychologically, some of us want to hear the grinding of chain on gears sooner. But the evolution to Hyperglide and more speeds on the rear happened very fast and a lot of it was due to pulley float along with ramps/teeth cutouts. The actual shifting these days occurs when the chain, under a lateral load by the top pulley reaches a release tooth (lower than most) and has a ramp that guides the chain to the next cog. We've all had that experience where we just missed that half rotation, and the crank slips a 1/4 turn or so and then we engage the next gear on a shift because we had to wait for the rear cassette to rotate a half turn over. So unlike the old days of heavy wear and tear requiring that we lighten our pedal stroke during shifts, a lot of riders are grinding hard and shift anyway, under load. I don't because I've grown up on the old school under friction shifting. But today's shifting has changed the way many newer riders approach shifting.

And I believe this is why Di2 has floating pulleys. Because it's going to shift for you and it leverages release locations on the cassette in conjunction with floating top pulley to allow the shift to happen under relative load without the need to lighten up much, but the cost of course is a slight amount of latency to get the shift lined up with cog tooth, ramp, and pulley ready. For some of us, I'm sure, this is bad because it seems less responsive. But I think that there is little advantage, even with top riders, to them shifting under load manually without pulley float. It only risks higher wear and lateral forces on the pulley and creates greater opportunity for misadjustment which would suck in the middle of a ride.

And one more thing that floating pulleys fix. Big riders with power going up a hill. Not all frames flex and reduce cable tension. But many do. And that's enough to really grind the gears and make noise.

In the end, the real test is how long it takes from initiation of shift to completion and how precise the adjustment has to be. I think Shimano's engineers have tested this and that's why they continue to have floating RD top pulleys. We could argue how the float might make shift initiation that much slower or reduce later load to applied to initiate a shift, but I think the design center around shifting has changed for all makers and we no longer truly derail chains so much as synchronize the shift with a location and ramp on the rear cog and allow the chain, with minimal side load, to be released to catch an adjacent cog. In this different design center, we don't need much side load, and we don't need to force the chain off one cog by rapid, forceful action. We just let it happen when the cogs line up twice per revolution in the rear for that to happen.

It's sort of like PCs and DRAM. We used to have chipsets with memory management that would page fault on a physical address request when it wasn't found in cache, and we'd randomly fetch that page. Then we evolved to have memory pre-fetchers, and now we have a synchronous bus that fetches memory every multiple clock cycles and only fetch pages memory, like every 12 or 15 ticks of the CPU. If you miss, then you gotta wait another 15 ticks or so. But overall, the speed of computers continues to go up. I would make the floating pulley akin to level zero cache. It holds data and instructions for the CPU and feeds the CPUs allowing it to keep pipelines full, meanwhile, page fetching occurs on a regular cycle to fill the cache but the cache keeps up with the CPU and keeps feeding it regardless. Question we're sort of asking here is if the cache has to be perfectly aligned to the CPU. And I ask how many cores does the CPU have and do they share that cache? We discussed all the possible things why Shimano put float into their top RD pulleys. Well, if multiple cores require coherent cache to pull data and instructions from, then it would be good to implement float so we don't thrash memory and TLBs and stall processors. Float is sometimes good.