You're wrong on multiple levels.
1. Generating random values for x^2, y^2, z^2 and taking their square root will only give you values in the x,y,z > 0 octant.
(But let's say you "fix" this by randomly multiplying them with -1.)
2. Taking the square root of a uniformly distributed random variable is no longer uniformly distributed.
3. Randomly reordering the coordinates won't fix your bias.
Here's a demonstration in 2D: https://jsfiddle.net/tz85wnxy/59/
Uncomment line 17, 18, 20 to see how it's still not uniform even if you randomly multiply the coordinates by -1 and reorder them.
Same process but minor tweak.
- if you flip the x-coord of a point near S, e.g. (-.9,0), you still get a point near S, e.g. (.9,0)
- if you swap the x-coord and y-coord of a point near S you still get a point near S.
- etc.
Sqrt(x^2 + y^2 +z^2) <= 1 reduces to x^2 + y^2 +z^2 <= 1
Generate the random value for x^2, any random value <= 1
Now we have y^2 +z^2 <= 1-x^2
Generate a random value for y^2, any random value <= 1-x^2
Finally generate a random value for z^2, any random value <= 1-x^2-y^2
Take the square roots of each number we generated to get the final coordinates.
But wait! Theres another step..the most important one! There are 6 orderings of (x,y,z) We must choose a random one because otherwise x will be biased towards higher values compared to y and z.
So I kind of lied. All we needed was 3 random values via the subtraction methods above. Then we need to randomly choose their ordering as (#,#,#)
This seems like the easiest most trivial solution that would be uniformly distributed.
Let me know if I am wrong.