FXAA is good in that it's really easy to implement and it's really cheap, the quality is not great. It has limited information to work with, and is completely incapable of handling temporal issues due to lack of sub-pixel information. If you use it you definitely want to do as ic0de recommends and grab the shader code and insert it into your post-processing chain as opposed to letting the driver do it, so that you can avoid applying it to things like text and UI. There's also MLAA which has similar benefits and problems.
You are correct that the "running the shader per-pixel" bit of MSAA only works for writing out your G-Buffer. The trick is to use some method of figuring out which pixels actually have different G-Buffer values in them, and then apply per-sample lighting only to those pixels while applying per-pixel lighting to the rest. For deferred renderers that use fragment shaders and lighting volumes, the typical way to do this is to generate a stencil mask and draw each light twice: once with a fragment shader that uses per-pixel lighting, and once with a fragment shader that uses per-sample lighting. For tiled compute shader deferred renderers you can instead "bucket" per-sample pixels into a list that you build in thread group shared memory, and handle them separately after shading the first sample of all pixels.
I also wrote quite a bit about this in the deferred rendering chapter of the book that I worked on, and wrote some companion samples that you can find on CodePlex.
Deferred lighting, AKA light pre-pass is basically dead at this point. It's only really useful if you want to avoid using multiple render targets, which was desirable on a particular current-gen console. If MRT isn't an issue then it will only make things worse for you, especially with regards to MSAA.
TXAA is just an extension of MSAA, so you need to get MSAA working before considering a similar approach. Same with SMAA, which basically combines MSAA and MLAA.
Forward rendering is actually making a big comeback in the form of "Forward+", which is essentially a modern variant of light indexed deferred rendering. Basically you use a compute shader to write out a list of lights that affect each screen-space tile (usually 16x16 pixels or so) and then during your forward rendering pass each pixel walks the list and applies each light. When you do this MSAA still works the way it's supposed to, at least for the main rendering pass. If you search around you'll find some info and some sample code.
As for the G-Buffer, as small as you can make it is still the rule of thumb. In general some packing/unpacking shader code is worth being able to use a smaller texture format. Reconstructing position from depth is absolutely the way to go, since it lets you save 3 G-Buffer channels. Storing position in a G-Buffer can also give you precision problems, unless you go for full 32-bit floats.