Fix CI on develop after #5585

[nlohmann]

The Ubuntu workflow on develop fails since #5585 (run https://github.com/nlohmann/json/actions/runs/37530611400). This PR fixes the three causes.

• ci_test_gcc, ci_test_standards_gcc (14): the new test test-diagnostics-optimized is built with -O3. At that level, the -Winline and -Wsuggest-attribute=pure warnings in the ci_test_gcc flag set fire on the library's inline functions. The test does not check these warnings, so they are now turned off for this test, together with -Wsuggest-attribute=const. The -O3 -Werror=array-bounds check for Release build fails when JSON diagnostics enabled #5742 stays. Reproduced with GCC 16.2 in gcc:latest; the test now compiles cleanly.

• ci_test_clang, ci_test_standards_clang (11): the static table in the test's to_json triggers -Wexit-time-destructors. The warning is now suppressed with DOCTEST_CLANG_SUPPRESS_WARNING, as other tests already do.

• ci_infer: 7 false NULLPTR_DEREFERENCE findings in tests/src/unit-pointer_access.cpp, caused by Create a value before giving it its type #5585. basic_json::replace_value() calls destroy() and then assert_invariant(). Together they create more paths than Pulse explores by default (pulse-max-disjuncts, 20). Pulse then loses track of the stored type and assumes get_ptr() can return null for a value it just constructed. Removing either call made the findings go away, but the library code stays as is. Instead, .inferconfig raises pulse-max-disjuncts to 40. Checked with Infer 1.3.0 (the CI version) on test-pointer_access: 7 findings with the default, none with 40.

• ci_infer: USE_AFTER_DELETE at include/nlohmann/ordered_map.hpp:262 (erase(first, last)). This is a false positive: Infer models std::string as keeping the buffer of a moved-from string, so the loop that destroys and re-creates elements looks like it destroys a buffer twice. It now has an @infer-ignore USE_AFTER_DELETE comment, the form already used in json.hpp and binary_reader.hpp. A full Infer run with pulse-max-disjuncts 40 and without this comment reported only this finding.

Two more failures came in with that merge, and both also affect develop:

• ci_test_standards_clang (14, libstdcxx), ci_static_analysis_clang (ci_test_clang): with JSON_NOEXCEPTION, JSON_THROW becomes std::abort(), so the parameters of binary_writer::throw_on_discarded() (Throw type_error.321 when serializing discarded values to binary formats #5761) are unused and -Wunused-parameter breaks test-disabled_exceptions. They are now cast to void. Develop had not shown this yet because these jobs were cancelled there.

• ci_cmake_options (ci_test_delete_deprecated_functions): the sax_parse cannot read CBOR that contains tags, including binary values with a subtype written by to_cbor #5676 regression test from Allow SAX parsing of tagged CBOR #5740 calls the deprecated sax_parse(span_input_adapter&&, ...), which JSON_DELETE_DEPRECATED_FUNCTIONS deletes. Those two checks are now skipped in that configuration, using the JSON_TEST_DEPRECATED_FUNCTIONS_DELETED pattern from unit-cbor.cpp.

• ci_clang_tidy: this job had been cancelled on develop, so its findings had piled up:

  • unit-diagnostics-optimized.cpp (this PR): to_json read past the end of its table for a value it did not contain. It now falls back to the first entry, as NLOHMANN_JSON_SERIALIZE_ENUM does. The test still fails with -Werror=array-bounds on the headers from before Create a value before giving it its type #5585.
  • unit-regression2.cpp (Make basic_json destruction allocation-free and non-recursive #5762): const/auto for the destroy() test values, and a NOLINT for the copy that check_destroy_edge_case() makes on purpose.
  • unit-serialization.cpp (Stricter fuzzer checks and boundary-value tests for buffers #5774): the expected strings are built with += instead of chained operator+.

  clang-tidy 22.1.8 (silkeh/clang:22, the CI image) was run locally on every test source and now reports nothing.

On the previous push (2e4d840), ci_infer passed ("No issues found"). It took 81 minutes, against 50 on develop. A local run with the same settings took 52 minutes, so most of the difference is probably runner variance.

Develop is merged in (#5740, #5766, #5774-#5777); none of those commits touch the files changed here.

Public API

No breaking changes. The library changes are a comment in ordered_map.hpp and two static_cast<void> statements in a private function of binary_writer; the rest is CI configuration and tests.

---

Written by Claude Code.

🤖 Generated with Claude Code

[gregmarr]

I thought that was fixed now.

[nlohmann]

It is fixed. The reordering in #5585 removed the false positive, so the comment was out of date. The test is now the regression test that keeps it fixed, and #5784 rewords the comment to say that.

Posted by Claude Code on behalf of @nlohmann.