A: Direct assemblage of matrices circumvents the necessity of writing to disk the EMAT and EROT files. Currently (at 5.6), all thermal elements for steady-state and transient analyses (linear and nonlinear) support direct assembly.
    For structural applications, only linear static and transient analyses support direct assembly for 2, 45, 92, and 95 elements. If your analysis is nonlinear, EMAT and EROT files are automatically written. I have attached a simple input file "solid95.inp" which demonstrates this -- run it with or without NLGEOM,ON to see the difference.
    Use of SURF154 does not create EMAT/EROT files when used with 92, 95, etc., either. The attached "surf154.inp" input file demonstrates this. I do not know all of the elements which can be used w/ 92/95/etc without generating EMAT or EROT files, although 154 is one I use quite often.
    As a result, I think that the reason why EMAT and EROT are written are because of the fact that contact or pretension elements trigger nonlinear equil iterations which, in turn, is not supported by direct assembly.

A: Although I am not sure why one would run 92 models w/o midside nodes (constant strain tets), the reason why the run times are not affected much by this change is because of zeros stored in the element matrix. It is similar to performance of degenerate tet 95s vs. 92s (please see XANSYS archives on this topic, subject is "SLOW HEX AND TET RUNS" on 5/17/00 which I posted) -- any degenerate element seems to still store additional information for its non-degenerate nodes.

A: As mentioned above, SURF154 and the other elements PLANE2, SOLID45, and SOLID95 are ones I can think of off of the top of my head. I do not know if any others are planned, although I hope direct assembly will be supported by most, if not all, elements. Support for other non-CG solvers would be nice, too.
    I imagine that the reason why EMAT is written for nonlinear structural runs is that with NLGEOM,ON (or contact present), the stiffness changes, so EMAT needs to be written/updated each step. However, not being a programmer and not being familiar w/ technical details, it seems to me that EMAT/EROT could be written in memory rather than to disk, so it doesn't seem that it would be necesary to write these files.
    There is an undocumented option DIRECT,ON you can play with to force direct assembly of equations for other elements, including contact. I would *strongly* recommend that you use this only to play with, not for production work, since this has not been fully tested/supported (unless someone else can recommend otherwise). With limited testing I've done, this seems to yield correct results compared w/ default, but, again, I am only mentioning this undocumented option for those who are interested in trying this just for kicks since I myself have not done extensive testing with this command.
    As a side note, ESAV (and OSAV) will be written for nonlinear analyses since element saved data (like strain history) needs to be retained in any nonlinear run for the next substep.