I wasn't part of that effort directly so I only saw what occurred at a slight distance but I'll suggest that that might not be a fair characterization. The SR2 team got quite a lot of direct support from the WinCE for Dreamcast team since it was a launch title and issues from them were given high priority. Current builds of SR2 were available at MS for testing continuously; I recall one of the early techniques the SR2 team used for the dust kicked up by the cars looked nice but absolutely tanked the framerate and they had to be advised to use another approach. They were certainly not left to stumble around in the dark.

As for performance, having an OS instead of coding to the bare metal of course imposes overhead but, from what I recall hearing, 3D performance of the WinCE SDK was on a par with Sega's SDK. The graphics engineers had access to Sega's SDK so they were able to do their own comparative benchmarking and they knew game developers would be scrutinizing that.

Ultimately, SR2's frame rate was entirely in the hands of the Sega Rally 2 development team by what code they used and what level of detail they tried to show. The cheat code for SR2 for the Dreamcast that reduces detail to increase framerate is an example of that. They were highly motivated to replicate the arcade experience as closely as possible and constrained by a need to hit the launch date and made choices that resulted in the frame rate issues SR2 has. Had they been using Sega's bare metal SDK, they may have made the same choices favoring arcade fidelity over stable framerate and therefore laying all the blame on WinCE isn't entirely reasonable.