Close 0002: guard the expected-lite selection, run reflected JSON iteration under ASan

[rmorozov]

Requested by Ruslan · project thread

Before: task 0002 reported a segfault when iterating an nlohmann::json array passed through Reflect, with no root cause. No test iterated reflected JSON arrays, and no sanitizer job built the nlohmann bindings. A consumer compiled at C++23 without the CMake/pkg-config flag could silently get std::expected as jinja2::Result<T> while the library uses nonstd::expected.

After: the cause is identified and its remaining instance is closed. template.h defaults expected-lite to its own nonstd::expected and stops with #error when std::expected was already selected in the translation unit. A new test covers iterating reflected arrays, and it runs under ASan+UBSan in CI.

Root cause. This was not a lifetime bug. Before 2.0 the library was built as C++14, so optional-lite and variant-lite used their own types. The probe driver was C++17, where the same headers alias std::optional/std::variant. CreateEnumerator() returns optional<ListEnumeratorPtr>, which the header-only nlohmann accessor and the compiled GenericListAdapter then read with two different layouts. I reproduced it on f2c05c0 by linking a C++17 test object against the C++14 library: UBSan reports load of value 64, which is not a valid value for type 'bool' in nonstd/optional.hpp and every loop renders empty. When both sides use the same standard the test is clean, on f2c05c0 and on master. 0071 (std vocabulary types, C++17 floor) removed the cause. Only the expected selection was still left open.

How.

• include/jinja2cpp/template.h: #ifndef nsel_CONFIG_SELECT_EXPECTED sets it to NONSTD, then the include, then #if nsel_USES_STD_EXPECTED triggers the #error. Builds through the CMake target or pkg-config already pass NONSTD and are unaffected.
• test/binding/nlohmann_json_binding_test.cpp: NlohmannJsonTest.ArrayIteration covers top-level, nested, if-filtered and filter-consumed (join, map) iteration. Its output matches Python Jinja2 on the same data.
• .github/workflows/linux-build.yml: the sanitizer matrix gets a bindings axis. The clang-18 address+undefined job now builds with nlohmann, the gcc-14 ASan and TSan jobs stay on boost, and job names are unchanged.
• docs/tasks/0002: marked done with the root cause.

Agent notes. Roles: verifier (1 round, no blocking findings). I applied its one nit: the test now uses o['values'], because in Python o.values is dict.values. Red pushes so far: 0. Checks run locally: full suite under clang-18 ASan+UBSan with nlohmann bindings, scripts/preflight.sh, and a gcc-14/clang-18 C++17/C++23 syntax matrix for the header guard. MSVC and the shared-library build were not run here; CI covers them.

🤖 Generated with Claude Code

https://claude.ai/code/session_01URYuQSkLMNwHqGAcDqQZjq

---

Generated by Claude Code