This sounds like regular snapping and it's something you can already do in Blender. You need to make a set of pieces that fit together at their border. You decide on the dimension of the pieces, i.e. 2x2 Blender units and then you place them in a grid as if they were tiles. You can also snap one object to another object's vertices to get a perfect fit (presuming the pieces were made to fit)

Blender won't restrict you when rotating and placing objects unless you tell it to. While the constraints to have a perfect grid are always there, the grid also needs to be in the artist's head so he/she knows where stuff should be placed.

This is something I've always wanted (big suprise) and no, it's nothing like regular snapping.

The idea is something more like a drag-and-drop constraints interface. Specific types of parts can go in specific positions with angular constraints, something of an implicit skeleton.

A grid is fine if you want to assemble a thing out of regular size tiles, not much use if you want to assemble a mech out of variable size chassis, arms, legs, weapons, etc.

Blender doesn't even support interactive grid-snapping. What it has is increment snapping so you can snap an object to two increments to the left of its current position, but not to absolute grid possition (-2, 0, 0).

Well with the 3D cursor you can do some pretty easy snapping in Blender, but otherwise something like this should be left to the actual game editor, no? I don't think Blender makes for a very good level-editor at all, and since the main point of having these repeated parts is to have the rendering offloaded to the GPU via VBOs, the engine needs to know of the individual parts, so a combined "level" mesh exported from Blender doesn't make much sense.

Acually they more or less abandoned the BGE, and the latest plan is to properly integrate parts of it with the main program as a sort of "interactive" rendering mode.

Besides, a good level editor (for most games, but not all) is actually more like a CAD program or at least something like sketchup, e.g. a modeling application that is geared towards architectural modeling. Blender is really not going in that direction at all, nor do I think it really should, because it is already too complex and full of feature creep.

Two projects that are worth to be mentioned here:

http://www.maratis3d.org & http://code.google.com/p/gamekit/

Both are meant to be full engines, but both also work well together with Blender.

GameKit can (is meant to) read directly from blender files, and maratis has a pretty complex blender addon, that allows to set up materials etc in a way maratis can read them later on. 

If they will come with editors, and how useful these will be for more general purposes, will be seen.

For reference:

http://code.blender.org/index.php/2013/06/blender-roadmap-2-7-2-8-and-be...

I don't think a lot of people do architectural modeling in Blender, but rather use the nice rendering features to make pretty pictures and videos of models they are created in another program.

There is also a plugin that adds more architectural CAD like features (e.g. booleans and such AFAIK) to Blender, and you are right that technically there is obviously a lot of overlap between those two types of programs, but a CAD program really has a quite different work-flow, and that is difficult to replicate in Blender without basically changing the entire way it behaves (keyboard shortcuts, GUI, 3D window interaction etc.).