Like all other parts of the pipeline, rigging heavily depends on the quality of the model. While this is not a problem on clients with a dedicated staff, it is definitely a bottleneck to the lesser specialized clients. As such, this is the first factor I consider when accepting a job. It is not advisable telling the client you can't get a good deformation during the process of rigging. It comes off as passing the blame. So the ability to scrutinize a mesh before the project starts is extremely important. In addition, this give you credibility that you are knowledgeable on the process.
Here are some of checklist I run through whenever I rig a mesh:
Symmetrical Mesh: Determine if this is unintentional and intentional symmetry.
Unintentional symmetry: Generally, you'd know that it is unintentional symmetry if the model looks symmetrical but the number of vertices of the left and right side of the model is not equal. In addition, you can check symmetry by performing re-symmetry yourself. Cut the mesh in half, duplicate it and scale in negative -X. If there are any overlapping meshes, then you have an asymmetrical model. When you have an unintentional symmetry, you either return it to client for clean-up or you do it yourself, with permission. I personally prefer the latter for better scrutiny and less unnecessary turnarounds. Before I mislead you, asymmetry is not bad per se. It only becomes so when it is unintentional which causes problems later down the line (i.e. mirroring weights are not working properly).
Intentional symmetry: Asking if the model is an intentional symmetry allows you to quote your project appropriately. Say you gave a uniform quote to a standard rig you referred in your reel. You accepted the task but it turns out the mesh is not symmetrical, which means you have more or less twice the skinning work. In other words, you are being underpaid. Of course, you can renegotiate but this could have been avoided if you asked some preliminary questions. Most mesh with intentional symmetry are monsters or creatures usually having an arm larger than the other. Depending on the client, scanned humans are maintained for their asymmetry.
Double Sided Mesh: These are usually present on clothes and clothing accessories where it is modelled like it has thickness. There are cases where it is permissible like a jacket where you can see the front and back side. Most of the time, unless it's on the edges, double side polygons are a nuisance not only to rigging but also a waste on UV space.
A Pose or T Pose: Either initial arm pose is permissible. In most cases, the only reason to choose one over the other is preference.The consensus is the following.
On T-pose. Takes a significant lesser amount to set-up since you can place joints easily. Easier to place pole vectors. There is almost always a pinching on the shoulders and clavicle. For skinny cartoony characters, this is less a problem. For realistic muscular characters, corrective shapes are almost needed to maintain appeal. .
On A-pose. There are several variations of A-pose. One is depending on the angle of the shoulders. They can be at 45 or 30 degrees. The elbows are bent. The fingers are curled. Needless to say, the further different you are from a T-pose, the longer it takes to model. A-pose is easier to model and pose in appeal since it is almost on a typical pose (i.e. arms side down).
Why not model it with the arms on the side down? You certainly can! But given it is an extreme pose (the other extreme is when the arms side up), it will require more corrective shapes to maintain an appeal. In our other words, T-pose and A-pose is actually a compromise to the extreme poses.
Here's the good thing about this. You can model in one pose, rig it in another pose and deliver still in a different pose. For instance, your modeller models mesh in an A-pose (elbows bent or hands curled). For the sake of rigging, you modify it in T-pose. But you deliver it in an in an arms side-down pose.
Isn't this time-consuming (working in a different pose)? To some degree it is but it has its merits. A rigger may rig a T-pose for an 1 hour but can take up to 3 hours on an A-pose. Modifying an A-pose to T-pose may take 30 mins. So the rigger can save 1 hours and 30 minutes in the process.
How about initial poses for legs? Generally, whatever initial pose you have for your arms, the same goes for legs for the sake of consistency. The reason that there's not much debate on this matter is because legs do not move as wildly as the arms. Even if it does, it almost lasts only for a few seconds since the legs has to come back to the ground to maintain balance.
In some cases, the reason to choose an initial pose is based on character movement. For instance, if you have an old sorceress who fiddles her hand as if casting a spell whenever she talks, it might be practical for her to have an initial pose with shoulder and elbow bent forward.
In my practice, I just choose a typical A or T Pose regardless of character's movement. There's really not much different since I use corrective shapes in most cases.
a) Maya's Human IK works best on a T-Pose
b) Unity's Mecanim works best on a T-Pose
c) For cloth simulation, A-Pose is easier to handle.
d) To be crude, all initial poses require corrective shapes to maintain appeal on extreme poses. This means that either T or A Pose will not save you from all the dirty work.
Front should be front and back should be back. This usually happens when you work with files from different software. For instance, Cinema4D is Z-backward system while Maya is Z-forward system. This is not technically critical but is really annoying down the pipeline. This will take a toll when you have two characters. One is modelled from C4D and the other in Maya. (For this illustration, both are rig in Maya) This means what is right in a C4D character is left in Maya. This means if you have a script that relies in left and right nomenclature for rig maintenance, you'd have to accommodate both system. You just have to have a standard Z-facing system. Either backward or forward is permissible. They just have to be consistent.
If you are asking, why not model entirely Maya? That is reasonable but more often than not software have different strengths. What is two hours in one software can be just 30 minutes in another.