Add missing diagnostic-positions test coverage (lifetime, input adapters, SAX)

[nlohmann]

Summary

Stacked on #5474 (fix for #5417), this PR adds the coverage requested in #5420 to the (now merged) tests/src/unit-class_parser.cpp, inside a new TEST_CASE("diagnostic positions: value lifetime, input adapters, and SAX") guarded by #if JSON_DIAGNOSTIC_POSITIONS. Trailing-comma coverage under positions (item 4 of #5420) is already closed by #5417, so it is skipped here as instructed.

Covered, as characterization tests (regression protection for existing behavior, not new behavior):

• Value lifetime
  • Copy constructor copies start_pos()/end_pos() recursively (root and every nested value).
  • Move constructor resets the moved-from value's positions to npos; the destination keeps the original positions recursively.
  • swap() — see "Found while characterizing" below.
  • Mutating a parsed document (operator[] adding a key, push_back, erase) leaves the parent's own (now stale) span and existing siblings' positions untouched; newly-added values get npos.
• Input adapters
  • json::parse(std::wstring): positions count transcoded UTF-8 bytes, not wide characters (demonstrated with L"{\"a\":\"éé\"}", where end_pos() == 12 but ws.size() == 10).
  • BOM-prefixed input: root's start_pos() is 3, reflecting the skipped BOM.
  • std::istringstream, std::ifstream, and iterator-pair inputs all report fully consistent (non-npos) positions.
  • Binary formats (from_cbor, from_msgpack, from_ubjson, from_bson) always report npos, since binary_reader never sets a text position.
• User-driven SAX: a user-constructed json_sax_dom_parser (no lexer pointer, e.g. built directly instead of going through json::parse()) reports npos for every produced value, since every if (m_lexer_ref) guard in json_sax.hpp is skipped without one.

Found while characterizing (flagging for a maintainer, not fixed here)

basic_json::swap(reference) (include/nlohmann/json.hpp, around line 3540) and the friend swap(reference, reference) that forwards to it swap m_data.m_type/m_data.m_value but never swap start_position/end_position — unlike copy-assignment's operator=(basic_json) (around line 1291), which does swap them as part of its copy-and-swap implementation. Concretely:

json a = json::parse(R"({"a":1})");
json b = json::parse(R"([1,2,3,4,5])");
swap(a, b);
// a is now [1,2,3,4,5], b is now {"a":1} -- values did swap
// but a.start_pos()/a.end_pos() still describe the OLD `{"a":1}` span,
// and b.start_pos()/b.end_pos() still describe the OLD `[1,2,3,4,5]` span

This looks like an oversight/inconsistency rather than intended behavior. Per the scope of this PR (test-only, characterization, no include/ changes), it is not fixed here — the test only pins the current behavior, with a comment pointing at the exact lines, so a future fix (or a deliberate decision to keep it) shows up as an intentional, visible change rather than a silent regression.

Validation

Compiled and ran unit-class_parser.cpp offline with -DJSON_DIAGNOSTIC_POSITIONS=1 (new tests exercised, 10511 assertions passed) and without it (new TEST_CASE fully compiled out, 10299 assertions passed, unchanged from before). Also checked both variants compile warning-free under -Wall -Wextra -Wpedantic -Wshadow.

Breaking change?

No breaking changes. This is a test-only change; nothing under include/ was modified.

Fixes #5420

— opened by Claude Code on behalf of @nlohmann