fix(build): make clean remove the test artefacts, and stop hardcoding clang-only flags

[balbasty]

Agent:claude-fastfields-to-teeny

Closesfastfields/fastfields-lib#131.

Why

Two halves of the same trap. make test CXX=g++ in a directory that had been
built with clang did not measure anything — while looking green. fastfields/fastfields-lib#131 was filed
from #54, where exactly that produced a "both compilers pass" claim that was
not true; the same trap is in the way of every subsequent both-compiler gate,
so it is worth ten lines to remove it rather than working around it again.

What changed

1. clean now removes the test artefacts. It removed build/*.o and
build/*.so but not build/testobj/ or the build/test_* binaries, so a
compiler switch in the same directory re-ran the previous compiler's binaries.
The output was real — it was just the wrong build's, which is worse than a
failure. clean now depends on a clean-test that removes both, plus the *.d
dependency files next to the objects.

2. DIAGFLAGS is chosen by probing the compiler.CXXFLAGS/TESTFLAGS
hardcoded -ferror-limit=1 -ftemplate-backtrace-limit=0, which are clang
spellings; g++ rejects both, so make CXX=g++ failed on the flags before
compiling a single file. g++ takes -fmax-errors=1 (the direct equivalent) and
has no counterpart for the backtrace limit.

The choice comes from $(CXX) --version | grep -ci clang, not from
pattern-matching $(CXX): a g++ that is really AppleClang, or a plain cc,
both then get the right answer. $(CXX) is only a name.

IS_CLANG := $(shell$(CXX) --version 2>/dev/null | head -1 | grep -ci clang)ifeq ($(IS_CLANG),0)
DIAGFLAGS = -fmax-errors=1
elseDIAGFLAGS = -ferror-limit=1 -ftemplate-backtrace-limit=0
endif

Verification

By doing the thing that used to be impossible — a from-scratch build under each
compiler, in the same directory, each compiling for real:

$ make clean CXX=g++
rm -f ./build/*.o
rm -f ./build/testobj/*.o ./build/testobj/*.d
rm -f ./build/test_distance ./build/test_distance_mesh … ./build/test_splinc
$ make -j4 test CXX=g++

suite	clang++	g++
test_distance	2352	2352
test_distance_mesh	4622	4622
test_distance_spline	704	704
test_posdef	5092	5092
test_pushpull	326	326
test_reg_field	525	525
test_reg_flow	854	854
test_reg_op	186	186
test_resize	4130	4130
test_restrict	191	191
test_splinc	4577	4577

failures: 0 / PASSED on every line, both compilers. Before this change the
g++ column could not be produced at all.

Build-only change; no source touched.

---

Generated by Claude Code

[balbasty]

Compile counts for the verification above, since "it built" is the claim this PR exists to make checkable. Counted from the full build logs:

$ make clean CXX=clang++ && make -j4 test CXX=clang++
clang++ compile invocations: 30 exit=0
$ make clean CXX=g++ && make -j4 test CXX=g++
g++ compile invocations: 30 exit=0

30 is every module object plus every test-driver object, i.e. the whole tree,
under each compiler in turn. Sample g++ command line, showing the probe picked
-fmax-errors=1 rather than the clang spelling:

g++ -std=c++17 -O3 -fmax-errors=1 -DTNY_MAX_RANK=64 -DFF_TEST_SPARSE … -c -o build/testobj/distance.o distance.cpp

and the clean that made the second pass real:

rm -f ./build/*.so
rm -f ./build/*.o
rm -f ./build/testobj/*.o ./build/testobj/*.d
rm -f ./build/test_distance ./build/test_distance_mesh … ./build/test_splinc

Worth noting how this bit me while writing the companion PR: my first attempt at
the two-compiler run piped make through tail, so the compile lines scrolled
away, and one of the two passes silently did no work — I only caught it by
grepping the log for compiler invocations and finding zero. Counting them is the
check; the check-count table alone cannot distinguish a real build from a rerun.

---

Generated by Claude Code