Interesting Max... Your rants are completely justified, anyone who deals with MR knows that
I asked because I have tested the Unified Sampler a lot, and have tried to gather information from other peoples experiences also... Overall I think it is a good step forward, but the lack of oficial information on how it works (exactly) and what are the best practices when using it make things shady for users, who at first may think it is actually slower then the ond sampler.
Using 1 local sample for glossiness and shadows, as you described, for example is a great thing! I find that is a great way to get consistent quality all throught the image with low render times, but there is no mentioning of this what so ever in the oficial docs. Shameful. (for those who dont know any thing about this visit this link: http://elementalray.wordpress.com/2012/01/29/unified-sampling-redux/
At the same time I found a little bug in that workflow. When a lens shader is connected to the pass energy on reflections verys depending on how many local samples you have (check out this thread Ive created at Mental Images forum: http://forum.mentalimages.com/showthread.php?9103-Energy-Inconsistency-with-local-sampling
). What is worst, it looks like (from my experiments) that the gamma correction in the render region is implement through some sort of unexposed Lens Shader. So you get different results between the Render Region and the Render Preview when using this technique (assuming you have no Lens Shader in your pass).
The good: When you know all this the Unified Sampler really feels like a great tool. I found myself rendering some (simple) scenes (with not much light bouncing) around in Brute Force mode. If you have tried Brute Forcing GI in MentalRay with the old sampler you know that it just was NOT POSSIBLE. Thin geometry and glossy reflections also have a great improvement in quality.
1) I am no developer but it is clear that either the shaders or the lights lack modern optimizations found in other renderers. Explicit lights sample as good (bad) as any emissive geometry which is freaking ridiculous when you work with renderers like Arnold or Vray. The hottest spots of contrast for the sampler are glossy light reflections, if you cant do that right no great adaptive sampler is going to save you.
2) The concept of only 1 quality control for the whole scene is a good. I just found the not linear way it works is very confusing. Example of an extreme situation: try having it at 0,005 your image sucks big time, have it at 0,01 your image looks a whole lot better; try increasing from 0,01 to 0,1 your image looks the same... even though the ratio between the first two values was of 2 and the second was of 10. Plus to me at 1 the image should be perferct, PERIOD. One should not have to find a random number, like 30 in your case, where your image is going to be good enough...
Other considerations about your scene setup:
1) Bucket size is waaaaay to small... this is very time consuming for adaptive samplers. That happends because they have to over sample the edges of the bucket since it does not know how good/bad the contrast is going to be on the samples from other buckets. So if your bucket size is 4x4 that means almost all pixels are being oversampled... Try a very large bucket size and see if differs a lot.
2) Reflection/Refraction depth is way too big... Speacially in very glossy reflections (your case) having more then 1 for trace depth is completaly unnecessary (visually). If I were you I would either bring them down to 1, and then increase it locally (in each material) as needed. Or the other way around...
3) Instead of going crazy with the quality (30) I would do things. First, check if your max samples isnt keeping you from reaching maximum quality... If you have a very high quality already (say, 5) and you still get noise in some materials, just increase the samples of that one material locally and bring quality down (not ideal, but...). Such a high quality may be wasting a lot of samples in areas where they are not needed...
If you have time, and pacience to play with those values, please share your results
And sorry for the long (huge) post...