Is there a fundamental reason the compilation can't be optimized away?
For example, given a header file A.h, couldn't you precompile it and generate a bloom filter of all preprocessor tokens contained therein (except ones defined in files that A.h includes)? Then if B.h (which includes A.h) changes, see if any of the preprocessor macros that are defined at the point of inclusion are in A's bloom filter. If not, you can safely skip the recompilation of A.h. If so, you should probably change A.h anyway to directly include the definition of any macros that are affecting its compilation.
I'm not sure how it gets away with this and works as well as it does. Maybe there is a behind the scenes check that I don't know about? In any case, it might be a good framework to add the caching you describe.
It also seems like this would be a win even if only used for rarely-changing system headers. As shown in the slides, for small projects the lines of unchanging standard includes dwarf the project specific code. Wouldn't it be a big win to avoid parsing and compiling all of these?
To get any benefit from the use of precompiled headers at all, you need to include practically every single header in your project in your precompiled header. Doing this is the opposite of modular.... it tightly couples each of your N compilation units with each of the M modules (changing any one of the modules bundled in the pch will force you to recompile all compilation units using the pch.)