ljoy, you might also enjoy the Shaywitz book. (Huge warning: I am no expert, so what I say here is my best understanding of the state of the evidence these days, just based on my own reading!)
One of the really fascinating things the Yale group is doing is using MRI to show the neurological deficit which is dyslexia. They can actually show the part of the brain activated when most people read, where a sort of "automated recognition" of words is stored (please don't quote me on this explanation!). Dyslexics don't activate that part of the brain when they read, but instead scramble to cobble together input from all sorts of miscellaneous places not meant for the job. Some verbally-gifted people can pull this off pretty decently and read with great accuracy - but if they are dyslexic, they will probably never read very quickly.
A huge shock to me was the recent discovery that my DH - who reads and writes for a living - reads at the same speed silently as out loud. I was beyond stunned (I am a particularly fast reader).
The problem is that using the "wrong" part of the brain to read is hugely demanding and inefficient. DH admits that reading is tiring, and while he reads large volumes professionally every day, he doesn't tend to do it for fun. So much brain power is used just for trying to make out the text, there's that much less left for higher-level comprehension, analysis, and just plain thinking about it. DH is good proof that a high-functioning dyslexic (with enough memory power) can be successful, and undetectable (we never suspected a thing until after DD was diagnosed). But we're also starting to recognize the price he has paid throughout his life for reading and spelling being so unusually demanding of time and energy.
So, to your specific questions:
A dyslexic may always get more out of an audio than written text, because they free up huge brain resources for higher-level thinking about it. So I would always encourage the use of both written and audio, even if reading skills seem adequate.
With dyslexia remediation, some of the reading workload will shift to the part of the brain designed for it (check out their MRI images of pre-post - it's really quite amazing). So reading, even if it was already highly accurate, can still become less demanding, faster and more fluent. And there's more brain left available for other tasks.
It sounds like spelling is also a challenge; remediation could probably reduce some of that pain and the time it consumes.
And finally, according to DH, if you could actually decode rather than sight-read the world, all those ^$%^$ god and monster names in Percy Jackson wouldn't have caused him such excruciating misery when he read all ten books out loud to DD last year.
