I pushed a change to help diagnose this since I'm not getting anywhere near that high a value for PRE implying poor hash mixing.
Please mail the complete output of that test from a failing run when you get one. Something like:
::: Running :NUMERIC-TYPE-HASH-MIXER
Dumping SB-KERNEL::*NUMERIC-TYPE-HASHSET*
Mask=3FF, maxPSL=4
(0) (INTEGER -4611686018427387905 -2)
(8) (INTEGER * 0)
(8 9) (INTEGER -173 104)
(11) (DOUBLE-FLOAT * 1000000.0d0)
(19) (INTEGER -4611686018427387904 -3)
(22) (DOUBLE-FLOAT 2.0d0)
... etc