refactor(test): simplify GlobalsGuard and TempDir, and test the harness

[TimD1-bot]

Note

Authorship: the content below was drafted by Claude Opus 5 (an AI coding agent) and
filed via gh under @TimD1-bot, a bot account operated by @TimD1. It reflects the
agent's analysis, not a statement authored by @TimD1.

Closes #119. Part of #45; follow-up to #118, which landed the unit-test harness.

Three cleanups to the unit-test scaffolding, done before the per-module test PRs build much on top of it.

1. Globals::VERSION/PROGRAM are now static

VERSION and PROGRAM are compile-time constants identical in every Globals instance, so they were never per-object state — but as non-static const members they deleted the implicitly-generated copy-assignment operator. Making them static const std::string (declared in src/globals.h, defined in src/globals.cpp) restores copy-assignment.

Every existing use goes through .data(), which is unchanged on a static const std::string, so there are no call-site edits: src/globals.cpp (print_version, print_usage), src/print.cpp (write_params), and the WARN/INFO/ERROR macros in src/defs.h.

2. GlobalsGuard no longer enumerates fields

With Globals copy-assignable, the destructor collapses from a 33-field assignment list to g = this->saved;. The public interface is unchanged — GlobalsGuard guard; at the top of a test behaves exactly as before.

The point is that a newly added Globals field can no longer silently escape restoration. Previously it would leak into the next test, and because gtest orders suites across translation units by link order, the resulting failure could surface in an unrelated test file.

3. TempDir drops the vector<char> copy

mkdtemp mutates its argument, which used to require copying the template through a std::vector<char>. Non-const std::string::data() is available in C++17 and the build already uses -std=c++17, so mkdtemp(tmpl.data()) works directly. The _XXXXXX suffix is untouched — mkdtemp requires exactly six trailing X characters.

4. tests/unit/src/test_harness.cpp

The scaffolding itself had no committed test, so a defect in it surfaced inside somebody else's test file. Eleven new cases, one gtest suite per helper:

Suite	Case	Covers
TempDir	CreatesWritable	the directory exists and a file can be written into it
TempDir	DistinctPaths	two live instances get different paths
TempDir	RemovesRecursively	a directory containing a file is gone after destruction
TempDir	PathJoin	path(name) is path() + "/" + name, with no doubled separator
GlobalsGuard	RestoresScalar	a mutated g.sv_threshold is restored at scope exit
GlobalsGuard	RestoresContainer	same for g.filters, which a field-by-field restore is likeliest to miss
GlobalsGuard	SilencesVerbosity	g.verbosity == 0 inside the guard's scope
WriteTmpVcf	ParsesUnderHtslib	the emitted file is accepted by parse_variants, the helper's entire contract
MakeFasta	SequenceReadable	sequence round-trips, uppercasing is applied, contigs stay independent
MakeCtgVariants	Roundtrip	count, order, and per-variant fields survive the builder
AllocReachOffs	SizeAndInit	the buffer is MATS*(max(x,o+e)+1)*(qlen+tlen-1) entries, all -2

test_harness.o is added to OBJS in tests/unit/build/Makefile; the existing pattern rule needs no change.

Verification

• cd tests/unit/build && make — builds clean (the one warning, dist.cpp:1070 unused parameter thread2, predates this branch).
• ./test_vcfdist — 22 tests from 9 test suites ran. / [ PASSED ] 22 tests., up from 11.
• cd src && make clean && make — links vcfdist; ./vcfdist --version prints vcfdist v3.0.0-b0, exercising both static members through .data().
• cd src && doxygen Doxyfile 2>&1 | grep -i warning — silent.
• cd tests && pytest — 3 passed.

[TimD1-bot]

Claude Opus 5 🤖: Rebased onto dev and fixed #128 in the same push.

Rebase

No conflicts. dev moved forward by #116 (per-variant provenance vectors) and #127 (push policy); neither touches a file this branch modifies, so the single commit replayed cleanly and nothing needed reconciling.

#128 — make_ctgVariants fixed

#116 inserted rec_idx, alt_idx and ploidy into the middle of add_var's signature, so make_ctgVariants passing var.supercluster eleventh bound it to rec_idx, and supercluster silently took its -1 default. That is exactly what failed MakeCtgVariants.Roundtrip here.

Took the richer fix #128 suggested rather than the minimal -1, -1, 0 literals: var_desc gains rec_idx = -1, alt_idx = -1 and ploidy = 0, and all four are passed explicitly — so the sibling test PRs can drive the new vectors rather than only work around them.

One deliberate deviation. The new fields are appended after loc, not inserted beside supercluster where add_var's own parameter order would put them. var_desc is aggregate-initialized positionally, so inserting mid-struct is the same silent-rebinding hazard #128 is about: it would quietly reinterpret every brace-initializer in the test PRs currently open. Appending keeps existing initializers bound to the same members, and the struct's doc comment now states the rule.

New regression test

MakeCtgVariants.ProvenanceFieldsReachTheirOwnVectors gives phase_set, rec_idx, alt_idx, ploidy and supercluster five distinct values and asserts each lands in its own vector, then checks that an unspecified var_desc still yields add_var's -1/-1/0/-1 sentinels. Confirmed it pins the defect: reverting only the make_ctgVariants call fails it on ploidies[0] and superclusters[0], alongside Roundtrip.

Also, #135's documentation half

var_desc::qual's comment claimed "add_var clamps to g.max_qual", which holds for var_qual only. It now reads Sets var_qual (clamped to g.max_qual) and gt_qual (unclamped). src/variant.cpp is untouched — whether gt_qual should also be clamped is #135's decision, and it changes published values.

Verification

• cd tests/unit/build && make — clean; the one warning, dist.cpp:1070 unused parameter thread2, predates this branch.
• ./test_vcfdist — 31 tests from 11 test suites ran. / [ PASSED ] 31 tests. Up from 22: feat(variant): add per-variant provenance and ploidy vectors #116 brought 8 ProvenanceVectors* cases in on dev, and this adds 1.
• ./test_vcfdist --gtest_filter='MakeCtgVariants.*' — both cases pass.
• cd src && make — links vcfdist; ./vcfdist --version prints vcfdist v3.0.0-b0, exercising both static members.
• cd tests && pytest — 3 passed.
• cd src && doxygen Doxyfile 2>&1 | grep -i warning — silent.

No src/ change beyond what the rebase brought in, so there is no count impact to report.