NeXT TurboColor panics in esp_read_reg() with any SCSI devices connected

I have a NeXT TurboColor with system ROM 3.1 v71.  It works perfectly
in OpenStep 4.2 (of course).  I attempted to netboot NetBSD, and I
finally got it to boot, but the kernel panics during the esp
initialization, before it attempts to mount root or start init, etc. 
Unfortunately I don't have the proper serial cable but I'll include
the output of "trace" from the debugger and part of the dmesg.

I have tried both the 2.0 release and the 20041106 snapshot; both
panic at the same point.  I also tried varying amounts of memory
installed (128MB, 64MB, 32MB), and aside from changing the addresses
of the kernel stack, it has no effect.

I also tried various SCSI configurations, to no avail.  My normal SCSI
config is one internal hard disk (Seagate ST51080N, 1GB) with OpenStep
4.2 installed (which NetBSD detects as sd0).  I have also tried an old
80MB IBM/Apple SCSI disk from a mac68k machine (not detected by
NetBSD); my normal disk plus an external CD-ROM (sd0 found, no mention
of CD-ROM); only a CD-ROM (no SCSI devices enumerated); no SCSI
devices at all (fails system SCSI test, presumably due to lack of
termination on the bus); and with the hard drive connected to the SCSI
bus but its power connector unplugged (this actually works and
continues booting).  All configurations except the last result in a
panic at the same point.

The messages from a panic (with only sd0, and 32MB RAM) are as follows:
===
... various dmesg stuff, typed by hand (argh!)...
Lookis like a MCS1850 (new style) clock chip.
scsibus0: waiting 2 seconds for devices to settle...
sd0 at scsibus0 target 1 lun 0: <SEAGATE, ST51080N, 0943> disk fixed