Fix the build: measure.c, the Makefile targets, and an LTO type mismatch

[Roxxik]

Four unrelated build defects, all pre-existing, found while trying to benchmark
the simulator. Reproduced on GCC 16.1.1 / GNU Make 4.4. Self-contained, it
touches no simulator code beyond one declaration in cbmbasic/.

measure.c does not compile under the project's own flags

$ make -f Makefile.measure
measure.c: In function 'main':
measure.c:371:60: error: 'end_x' may be used uninitialized [-Werror=maybe-uninitialized]
measure.c:382:60: error: 'end_y' may be used uninitialized [-Werror=maybe-uninitialized]

The = 0 in

uint8_t end_a, end_x, end_y, end_s = 0, end_p = 0;

binds to end_s and end_p only. It is not a live bug, every read is preceded
by a write on the j == 0 pass, and the guarding condition depends only on k,
which is fixed by the enclosing loop. GCC cannot prove that, and which of the
three variables it names shifts with the optimisation level, so it can look
intermittent. This initialises all five rather than suppressing the warning.
Different opt levels shift this. On -O0 there is no warning.

Make targets named after directories that already exist

Makefile builds a target called cbmbasic, and cbmbasic/ is a directory.
Make stats the directory, whose mtime is bumped whenever an object file is
written into it, and declares the target up to date:

$ rm -f cbmbasic/cbmbasic
$ make
make: Nothing to be done for 'all'.

make clean && make always works and touching a source file does trigger a
relink, so this stays invisible day to day. It shows up after a failed link:
every subsequent make reports "Nothing to be done for 'all'", which reads as a
successful no-op, and a stale binary can sit there while make insists the tree is
current.

Makefile.apple1basic has the same collision, and there it is fatal, because it
links with -o apple1basic:

/usr/bin/ld: cannot open output file apple1basic: Is a directory

This names each target after the file it actually produces and marks the real
phony targets as such. .PHONY: cbmbasic would also stop the misfire, but by
forcing a relink on every invocation.

apple1basic is missing netlist_sim.o

Fixing the output path above uncovers a second defect the link error had been
masking: netlist_sim.o is absent from its OBJS, so the link fails with
undefined references to recalcNodeList, readNodes, isNodeHigh and
destroyNodesAndTransistors. That target cannot have linked since
netlist_sim.c was split out.

N/Z/C are declared unsigned char and defined int

cbmbasic/glue.h declares extern unsigned char V, B, D, I, C, N, Z; while
cbmbasic/runtime_init.c defines int N, Z, C;. runtime_init.c does not
include glue.h, so no single translation unit sees both and an ordinary build
cannot notice. Under -flto GCC compares declarations across translation units,
reports -Wlto-type-mismatch for all three, and -Werror turns that into a link
failure.

Every value stored in them is 0 or 1 - N = P >> 7, Z = (P >> 1) & 1,
C = P & 1, the SETZ/SETSZ/SETNC macros, and the plain C = 0 / C = 1
assignments in runtime.c - so this narrows the definition to match the header
rather than widening the header. That also matches how A, X, Y and S are
already declared and defined.

Verification

All three targets build clean from scratch, measure characterises all 256
opcodes, and deleting a binary triggers a relink. cbmbasic --benchmark output
is byte-for-byte identical to a pristine build of master, with and without
-flto: 33155 half-cycles, same final CPU state.

Notes

• lto does not improve runtime.
• -march=native is slower on my machine
• using PGO is a slight improvement, but only when done on the same workload. Profiling on cbmbasic and running on measure is worse. Same the other way around.

I used an LLM to help me out in this work, but I manually reviewed all changes made.