Check for mk and limbo before the macOS builds

[Ticed]

What this changes

build-macos-headless.sh and build-macos-sdl3.sh assumed mk and limbo
were already on PATH. Neither checked, so a missing tool surfaced as a bare
command not found from inside the build rather than as a statement of what to
do about it.

makemk.sh builds mk alone. limbo still needs libmath, libbio, lib9 and
libsec built after that, so "mk present, limbo absent" is the state a
half-finished bootstrap leaves behind, and it is the one that costs time: the
build gets several minutes in, through the whole emulator compile, before it
reaches limbo and dies.

Both scripts now refuse up front when either tool is missing, name which one,
and name the bootstrap.

Both, because the hole is identical in each and build-macos-sdl3.sh is the
one AGENTS.md lists first.

Tests

tests/host/build_macos_preflight_test.sh is new. For each script it covers
neither tool present, and mk present with limbo absent. It stubs
pkg-config so the SDL3 probe cannot decide the result, and a failed assertion
prints what it expected and what it got.

Against each script before this change:

neither present   build starts, fails inside mk
mk present        build starts, dies later at "limbo: command not found"

After:

build_macos_preflight: PASS

The exit status was already correct in both. set -e is set and each ends in an
explicit exit 1; a failing step has always returned non-zero. Only the
preflight was missing.

[pdfinn]

The premise is right — "mk present, limbo absent" is exactly the state a
half-finished bootstrap leaves, and dying minutes into the emulator compile
with limbo: command not found is a bad way to learn that. Two problems
though, and the first means the message sends people down a path that will not
work.

The bootstrap instructions are incomplete

Run ./makemk.sh, then build libmath, libbio, lib9, and libsec in order before
rebuilding limbo.

limbo/mkfile declares:

LIBS=	bio
	math
	sec
	mp
	9

YFILES=	limbo.y

So libmp is missing from the list, and limbo.y means utils/iyacc has to
be built first as well — Inferno's yacc, not the host's. Following the message
literally gets you to a failure a step later than before, which is worse than
no advice.

I built limbo from C on a clean tree last week while working on #559; the
order that works is:

SYSTARG=$(uname -s) OBJTYPE=<arch> ./makemk.sh
for d in lib9 libbio libmp libsec libmath utils/iyacc limbo; do
    (cd $d && mk install)
done

lib9 first rather than third, since everything else builds against it. That
whole sequence takes about 30 seconds, which is worth saying in the message —
people put off bootstraps that sound expensive.

The test fails on the machines the code is for

expect_refusal runs env PATH="$TMP" "$ROOT/$script", but the first thing
both scripts do is

export PATH="$ROOT/MacOSX/arm64/bin:$PATH"

so on any macOS box where the native tools are actually built — which is every
machine that has run these scripts before — mk and limbo are found, the
preflight check does not fire, the build proceeds, and expect_refusal reports

build-macos-headless.sh: unexpectedly succeeded without mk limbo

It passes here on Linux arm64 precisely because MacOSX/arm64/bin does not
exist:

build_macos_preflight: PASS

So the coverage is inverted: green where the code never runs, red on the
platform it was written for. That is worse than #565's problem, because it is a
false failure rather than a false pass — someone will hit it locally and
conclude the test is broken.

Neutralising $TMP is not enough while the script re-adds the real bin
directory. Simplest fix is to move $ROOT/MacOSX/arm64/bin/{mk,limbo} aside
for the duration and restore them in the existing trap, so the script's own
PATH line finds nothing either. Copying the tree would also work but costs
more.

The check itself

The loop and the message shape are good, and doing both scripts is right — the
hole is identical and build-macos-sdl3.sh is the one AGENTS.md lists first.

One note: command -v is the correct POSIX spelling, but both scripts are
#!/bin/bash and use [[ ]], so they are bash already. Not a problem, just
inconsistent with CONTRIBUTING.md:226 in the same way the rest of these
scripts are — not something to fix here.

Fix the library list and the test's PATH handling and this is good to go.

[Ticed]

Both fixed, and thank you for the second one — the test really was inverted.

The bootstrap recipe. Corrected to
lib9 libbio libmp libsec libmath utils/iyacc limbo, with lib9 first and a
note that the whole thing takes about half a minute. Two deviations from your
sequence, both of which I hit by running yours verbatim on a clean tree:

SYSTARG=$(uname -s) OBJTYPE=arm64 ./makemk.sh
  regexec.c:1:10: fatal error: 'lib9.h' file not found

makemk.sh builds against $ROOT/$SYSTARG/$OBJTYPE, and there is no
Darwin/arm64/include — mkfiles/ has only mkfile-MacOSX-arm64 and
mkhost-MacOSX. mkconfig maps Darwin to MacOSX for mk install, but
makemk.sh does not, so the message pins SYSTARG=MacOSX. And the loop needs
ROOT and MacOSX/arm64/bin on PATH or every iteration is
mk: command not found. The message now carries both, and the test greps for
them so a later tidy-up cannot quietly restore advice that does not work.

The test. Rather than move the real mk and limbo aside, it now copies
the two scripts into a fake tree with an empty MacOSX/arm64/bin and runs them
there — both derive ROOT from $0. That is your "copy the tree" alternative
at two files instead of a tree, and it avoids the failure mode the move-aside
version had: the trap cannot catch SIGKILL, so a CI timeout kill between
stashing and restoring would have left a developer's toolchain gone. kill -9
part-way through now leaves the binaries byte-identical.

Verified on macOS with real mk and limbo present: the pre-fix test fails
exactly as you described — the script sails past the preflight and dies later —
and the fixed one passes with the refusal message in the output rather than
inferred from an exit code. Both scripts, both cases.

Left the command -v / [[ ]] inconsistency alone as you asked.

[Ticed]

Rebased onto current master (df34b02). Force-pushed 4310307 -> 5a08c85.

git range-diff reports every commit unchanged — same content, same messages,
only the base moved. The five commits that landed underneath are #583, #582,
#566, #578 and #580; none of them touches this branch's files.

Re-verified on macOS after the rebase.

tests/host/build_macos_preflight_test.sh passes on this macOS host with real
mk and limbo present — the case that was inverted before. Both binaries are
byte-identical by sha256 before and after the run, so the fake-tree approach
still avoids the SIGKILL hazard the move-aside version had.

#578 fixed the ClusterFuzzLite link break, so Fuzz should now be green here
rather than red for a reason that was never this branch's.