TL;DR: cast time is what you'd expect ± 500 ticks (distributed linearly) in each direction, regardless of the original cast time percentage.
Running this with -dddd shows that all mages on the same side doing the same thing start casting at the same time, so there's no variance in starting.
If I sort my data (this is for 100% cast time) I get this deliciously linear thing. It appears jagged because the debug log only ever prints increments every 10 ticks, so everything in the data is a multiple of 10.
Then it was just a simple matter of checking minimum and max times the spell got cast at, which revealed two numbers exactly two apart.
Then I repeated the whole thing on the same spell modded to 125% cast time, which increased the casting time but not the distance between min and max values recorded - that was still 1000. Also not shown is that the distribution is still absolutely linear.