When you have a raw sensor file in the Bayer format of 12-14 (linear) bits per channel and non-standard color filter response that you want to display on your typical display/printer of 8bits (2.2 gamma) sRGB/ARGB (standardized primaries) device, there is going to be some "loss" somewhere along the chain nomatter what. There is also going to be a radical change in numerical values, and this is a good thing as it makes the resulting image worth looking at.
The argument about Lightroom being "non-destructive" is that it separates image data (pixels) and editing instructions. Just like Microsoft Word is different from a typewriter - it does not save an image file containing the pixels of the rendered text. Rather, it works on a set of instructions that can be used to render the text at any time, but also do spell-checking and swapping the font. The original raw file is of course never touched, but the neatness of Lightroom is that it will read that raw file into memory, do your edits, and render the result to screen/export/printer. If you change one parameter, it will re-do this process (it may do internal optimizations, but analyzed as a user-oriented black-box I think this is correct). The important "philosophical" implication is that Lightroom isolates user 1. intent ("make the image brighter") and 2. image processing ("multiply all pixels by 2"). If the Adobe guys at some point find more clever ways to transform 1->2, they can do so (and atleast with the demosaicing in LR3, they did). They may theoretically do limited-precision integer processing in one version of Lightroom, and convert to single-precision floating point in another version as long as they deem that it gives only better end-result without compromising the intent of the original edit (this is mind-bogling because some users might actually want editing-induced banding and quantization and consider it a part of their edit. They may not be pleased with a radical new engine that does the processing "better").
Are you having problems with the end-results in Lightroom today? I guess that if you went to dcraw or something similar, you would have access to the ideal documentation of image processing (open-source code) so that you could put your mind at rest, or improve parts that you did not like.