A/UX Kernel Switch 1.18 - switch the A/UX kernel on the nuCF card between the normal and the
diagnostic Carrera040 start code, and turn the slot cards fix on or off, from Mac OS,
without taking the card out.

Get it onto the Mac (either file):
  A-UX Kernel Switch.hqx   BinHex: decode with StuffIt Expander
  A-UX Kernel Switch.bin   MacBinary
Run it under Mac OS 7.5, extensions on or off.  It finds the nuCF card and the A/UX
kernel, shows what it has, and asks four things:
  1. Diagnostic / Normal / Quit     the start code
  2. Slot cards fix On / Off / Quit the fix that lets A/UX use expansion cards (the
     Ethernet card).  Keep it Off: with it on the scan hangs at the Ethernet card.
  3. Fixes added since 1.9 On / Off / Quit (see 1.17 below); keep them On
  4. Carrera mode MED / FAST / 68030: the mode stored in the kernel (launch -k overrides it)
It only writes after checking every place it changes holds a version it knows (the
kernels of 2026-11-09 to -11, 1.0-1.18), and reads the result back.

New in 1.18: the fourth question sets the Carrera mode stored in the kernel, so MED
needs no "launch -k" flag.  With MED (write-through) A/UX now reaches the Finder both from
the BlueSCSI launch disk and from 7.5.5 with extensions off; FAST (copyback) corrupted
memory: the Finder bus errors and the random hangs.  The kernel uses its stored mode also
when the Carrera extension started the 68040, which is why FAST hurt there too.
Recommended: Normal, slot cards fix Off, fixes since 1.9 On, MED.
Tested in the emulator: only the mode byte changes, and the app reads it back.

In 1.17: the slot-probe fix (1.6) is back in for good; the third question now covers
only the three later additions - the kernel's sExec (1.9), the clean-up at the nuCF
driver (1.11), the caches off for the slot scan (1.14).  1.16 froze at "Launching...":
the slot-probe fix is what lets the 68040 get through a probe fault in the slot scan, so
taking it out made the scan hang; that run said nothing about the Finder crash.
Tested in the emulator: from a 1.16 kernel with everything Off, 1.17 with Off puts the
probe fix back and leaves the other three out, and the BlueSCSI launch disk boots to the
A/UX Finder.
In 1.16: a third question, "Fixes added since 1.6" On / Off.  Off takes out all four
things I added to the kernel since 1.6 at once - the slot-probe fix (1.6), the kernel's
sExec (1.9), the clean-up at the nuCF driver (1.11), the caches off for the slot scan
(1.14) - and leaves the kernel otherwise as it is (start code, slot cards fix, the older
fixes 1-10).  On puts them back.  The "Done" message says which you have.
Tested in the emulator: Off, then the BlueSCSI launch disk -> A/UX Finder; On again gives
back exactly the 1.15 kernel.
In 1.15 (Normal and Diagnostic): the kernel's own sExec (1.9) only acts during the
boot-time slot scan.  The BlueSCSI launch disk test (7.0.1 + Carrera software, the card's
kernel) got A/UX up but the Finder bombed with a bus error, so the problem is in the
kernel as it is now, not in the 7.5.5 extensions-off start.  Checked: the kernel fixes
from the working images (1-10, slot probe, write-through, SCSI, MOVE16, the ROM
cached...) are all still in, and the Mac environment's System (7.5.3, BlockMove fix) is
the same as in the working images.  Of what I added since, 1.9's sExec is the one that
also ran after boot: every Slot Manager sExec, also the Mac environment's, went through
it, and its "is the block where the kernel put it" check assumed the boot-time memory
pool.  Now it lets every call outside the scan straight through.
Tested in the emulator: forcing the bad block pointer at the scan still gets the repair
(row 4 column 14) and A/UX comes up to the Finder (Diagnostic).
In 1.14 (Normal and Diagnostic, always written): the 68040's caches are off for the
slot scan.  What the 1.13 photo showed: 13 exceptions, exactly the emulator's (the ROM
probing slots D down to 1), the last at F1FFFFFF, then a hang with no further exception
and before the built-in video's start-up: so the hang is in the card start-up code the
scan runs next, not in the probing.
The ROM is a IIsi's (68030 only): its Slot Manager copies card code (sExec blocks,
drivers) into RAM and jumps to it without pushing the 68040's data cache, so with the
caches on the 68040 can execute stale memory instead of the code just copied - any
crash, hang or wild jump, varying from run to run (your $04000000 jumps, the scan not
coming back, 1.11 getting through and 1.12/1.13 not).  The emulator has no caches.
The kernel already turns the caches off for the scan on a 68030 but leaves them alone on
a 68040; at the scan they are on (the emulator shows CACR $80008000: both caches on).
1.14 pushes the caches and turns them off when the scan starts and leaves them off until
the kernel's own cache start-up (which prints "mc68040 caching enabled") turns them on
again.  The Diagnostic console line now also shows what CACR was:
  scan faults: 13, CACR at scan 80008000
In 1.13 (Diagnostic only): what the 1.12 photo showed: rows 1-3 complete, row 4
columns 0-3 (the slot scan started), then nothing, and the screen was never grayed, so it
hung inside the scan before the built-in video.  1.11 got through the scan and 1.12 only
added logging in the slot-probe handler, so either the scan doesn't always end the same
way on your machine or the logging changed it.  1.13 also draws the log on the screen, as
the scan goes, so a hang in the scan leaves it visible.  From line 268, left half, rows
of 32 cells 7 lines apart (a long cell = 1, a tick = 0), redrawn at every exception:
  1 how many exceptions so far
  2-5 this one: PC, format/vector, fault address, a2
  6-9 the one before: the same four
In the emulator the first is PC 40805E84, 7008, FDFFFFFF, a2 40805E8C, and each next one
has a fault address one slot lower (FC..., FB..., down to F1FFFFFF), 13 in all; then the
built-in video grays the screen.
In 1.12 (Diagnostic only): with 1.11 the kernel starts fully, but the Mac side
("Can't attach to the toolbox environment") fails, most likely because the slot scan
doesn't come back the normal way and leaves the Slot Manager's tables part-built.  1.12
records every exception the slot-probe handler takes during the scan (up to 32) and
prints them on the console right after the nuCF driver's lines, then waits about 20
seconds before going on:
  scan faults: N
  0: PC fmt/vec fault-address a2   1: PC fmt/vec fault-address a2
  ...
In the emulator there are 13, all the ROM probing the empty slots:
  0: 40805E84 7008 FDFFFFFF 40805E8C   1: 40805E84 7008 FCFFFFFF 40805E8C ... F1FFFFFF
Anything different on your machine is the lead.
In 1.11 (Normal and Diagnostic, always written): what the 1.10 photo showed.  At the
nuCF driver's start the 68040's transparent translation was still set the way the slot
scan sets it (DTT1 $FE00A040 = the built-in video only; DTT0 $403FA040 = the kernel's
start value) instead of the split setting the scan's exit code puts in ($400FA000 /
$50AFA040).  So slot D ($FDD00000) had no mapping at all (MMUSR 0) and the driver's
first read of the card hung.  The scan's exit code never ran: the scan came back some
other way, and the slot-probe exception vector the scan installs was most likely left
in place too.
1.11 checks at the start of the nuCF driver: if the scan's exception vector is still in,
it puts the kernel's back; if the scan's TT setting is still in, it runs the exit code.
If it had to do either, it prints one line (Diagnostic and Normal):
  scan exit: N vec X tt1 saved X X ret X X dtt0 X dtt1 X
  N = 1 vector put back, 2 TT put back, 3 both; vec = the vector found; dtt0/dtt1 =
  before.  It doesn't explain why the scan exited that way; that's the next question.
Tested in the emulator by forcing both leftovers before the driver starts: the line
appears, the driver finds the card and A/UX boots to the Finder.
In 1.10 (Diagnostic only): console lines through the nuCF driver's start-up, where
your 1.9 photo stopped (after "nuCF: slot 13; table ...", before the tape driver's line):
  nuCF: probe FDD00000 mmusr X dtt0 X dtt1 X sr X   about to read the card's status
  nuCF: probe -> 1                                 the card answered (0 = didn't)
  nuCF: identify                                   about to send IDENTIFY
  nuCF: identify -> 0                              done (-1 = the card timed out)
  nuCF: <model>, <size> MB, as SCSI ID 6 ...       (the driver's own line)
The last line you see says which step it is stuck in.  In the emulator:
  nuCF: probe FDD00000 mmusr 3 dtt0 400FA000 dtt1 50AFA040 sr 2704
In 1.9 (Normal and Diagnostic, always written): the kernel runs the Slot Manager's
sExec itself, checking the card start-up code block before jumping to it.  Your crash is
sExec jumping to $04000000: the block pointer it got back was $03FFFFF8 (the end of RAM),
not the block the kernel had just allocated for it.  Nothing in the 7.5.5 System's
leftovers that the emulator can reproduce explains it (the emulator with the same 7.5.5
boots), so this works around it instead:
  - block where the kernel put it: run it, as the ROM does
  - block pointer wrong: use the block the kernel just allocated, if it holds valid
    code (the copy may well have gone there), and run that.  A block at row 4 (line 190) column 14.
  - neither holds valid code: skip that card's start-up code and go on with the scan.
    A block at row 4 column 15.
  When either happens it also draws, from line 268 (cells as in the dump below):
    spBlock, spResult (the bad pointer), spsPointer, spSize, +$30 (slot/ID), pool end,
    pool length, the 2 longs at spResult (all ones if not RAM), the block it used.
  Tested in the emulator (7.5.5, extensions off): forcing spResult to $03FFFFF8 at that
  sExec, both ways boot to the A/UX Finder (Diagnostic, slot cards fix off); unforced,
  Diagnostic and Normal boot to the Finder with the fix in.  Whether it is enough on
  your machine depends on why the pointer is wrong, which this can't know: if more is
  corrupted than that one pointer, it will stop somewhere later instead.

The dump (from line 268): rows of 32 cells, 8 pixels each (a long cell = 1, a tick =
0), 7 lines apart, two columns:
  left  (x 0-255):   SR:format/vector, PC, EA, SSW:WB3S, fault address, a0, a1, a2,
                     the stack pointer before the error, the long on top of that stack
  right (x 256-511): the parameter block sExec's caller passed, its spResult,
                     spsPointer, spSize, +$30 (slot/ID), +$34, then the first 4 longs
                     of the block spResult points to (all ones if not in RAM)
In the emulator with the same crash forced: 0015CEE4 0030458C 4087ED36 0000047C
00002200 000F2600 02020000 00000004 4E56FFBC 317C0001.

Blocks are 16 pixels wide, 32 pixels apart, counted from the left edge.
Row 1 (line 100; the Carrera start): 0 kernel entry, 1 on the 68030, 2 handing over,
  3 the 68030 is back (gave up), 4 the 68040 is running, 5 entering the kernel,
  6 no Carrera, 7 68040 exception before kernel vectors, 14/15 jumpers.
Row 2 (line 130): 0 cpuinit 1 boardinit 2 memsize 3 mmusetup 4 vadrspace 5 (never)
  6 page tables 7 setup_040 8 turnon_mmu 9 setup_lowmem_globals 10 setup_autovectors
  11 mktables 12 reginit 13 patch_autovectors
Row 3 (line 160): 0 strinit 1 via1init 2 via2init 3 egret_init 4 iop_init 5 video_init
  col 15: an access error while the blocks were on screen; the register dump below.
Row 4 (line 190; inside video_init):
  0 initMacEnvironment   1 the ROM's init   2 patches done   3 slot scan starts
  4 slot scan returned   5 secondaryInit    6 secondaryInit returned
  7 video_find  8 getDevBase  9 getVPBlock  10 getVideoDriver  11 video interrupt
  12 video driver open  13 gray screen
Row 5 (line 215; during the slot scan, slot cards fix on): column = slot number
  ($9-$E) the scan was let through to (the card is there).
Row 6 (line 240): column = slot the scan probed and was refused (empty / nuCF), or
  column 0 = a fault that wasn't a slot access.
Expected on your machine: row 5 columns 11 ($B Ethernet) and 14 ($E video), row 6
columns 9, 10, 12, 13.  The scan grays the screen when it gets to the built-in video,
which wipes the blocks drawn so far, so blocks you can see mean it stopped before that.

In the emulator: all of the above to the Finder for Normal and Diagnostic, slot cards
fix on and off.
