Update doctest to 2.5.3 and fmt to 12.0.0

[martinus]

Updates the meson wrap subprojects to their newest packaged versions, which as a side effect restores the local clang build that #195 was filed about.

dependency	before	after	note
doctest	2.4.12	2.5.3
fmt	11.2.0	12.0.0	newest in wrapdb; upstream 12.2.0 has no meson packaging
abseil-cpp	20250814.1	unchanged	already the newest wrapdb has, and the dependency is commented out

Why this fixes #195, and why the fix there was the wrong one

doctest 2.4.12 does not build with clang 22 at all — 153 errors and no binary:

error: '__COUNTER__' is a C2y extension [-Werror,-Wc2y-extensions]

DOCTEST_ANONYMOUS uses __COUNTER__, C2y standardises it, and clang 22 warns under werror=true. That is why gcc has been the only compiler able to build this suite locally, and it is why the structured-binding bug in #192 reached CI instead of being caught before the push.

Worth recording: commit 794aa86 tried to solve this with include_type: 'system', and it never worked. A diagnostic about what a macro expands to is reported against the line that used the macro, so it counts as this project's code no matter how the header was included. I've corrected that comment rather than leaving it asserting something untrue.

The regression 2.5.3 brings, which is why this is not a one-line wrap bump

2.5.3 fixes the warning by using __LINE__ instead of __COUNTER__ on clang 22+. But __LINE__ is only unique within one file, so two test cases on the same line collide as soon as a unity build puts their files in one translation unit:

error: redefinition of 'DOCTEST_ANON_TMP_9ITERATOR<std::tuple<Type, Rest...>>'
  test/unit/insert.cpp:9  <-  previous definition at test/unit/insert_or_assign.cpp:9

Both files open with TEST_CASE_MAP on line 9. This is not a corner case:

• there is a Linux (gcc, C++17, unity) CI leg, and
• every lane of scripts/mutate/mutate.py is --unity=on — there the failure would surface as every mutant returning compiler instead of a real verdict, i.e. a mutation run that looks like it ran and tells you nothing.

The fix takes __COUNTER__ back via DOCTEST_COUNTER, which is doctest's own customisation point (it only picks a default #if !defined), and silences the warning doctest was avoiding. Both halves are required — the flag alone leaves doctest choosing __LINE__ (unity stays broken), and the define alone reintroduces the warning as an error. compiler.has_argument() keeps the clang-only flag off gcc; verified it is absent from gcc's compile_commands.json and present in clang's.

Verification

Every configuration built and the full suite run (539 cases, ~5.4M assertions):

configuration	result
clang C++17 / C++20 / C++23	pass
gcc C++17	pass
gcc unity (--unity-size=16, the CI leg)	pass
clang unity — broken before the fix	pass
gcc 32-bit (-m32, switches fmt to builtin)	pass
gcc asan+ubsan, clang asan+ubsan (CI's flags)	pass
libFuzzer fuzz_api / fuzz_string, corpus replay	pass
scripts/test_mutate.py (155 hermetic tests)	pass
end-to-end mutation run under clang unity lanes	caught

Two local failures were checked against the baseline and are pre-existing and environmental, not from this change:

• clang-tidy 22 flags 20 issues in unordered_dense.h, a file this PR does not touch — CI pins clang-tidy-18 precisely because check families grow between releases.
• clang-tidy-18 cannot parse gcc 16's libstdc++ (system_error), so the lint job is not locally reproducible either way.

libc++ is not installed on this machine, so that leg is left to CI.

Follow-up

Item 1 of #195 (add -Wno-c2y-extensions to restore the local clang build) is superseded by this PR — the flag is here, but on its own it would not have been enough, since the real blocker after the bump is the unity collision. Items 2 and 3 (self-calibrating estimates, incremental --json) are untouched and still stand.

🤖 Generated with Claude Code