Restore a duplicate key's prior value when the callback rejects its new value

[nlohmann]

Summary

In json_sax_dom_callback_parser::key(), when the callback accepts a key, the code does:

object_element = &(ref_stack.back()->m_data.m_value.object->operator[](val) = discarded);

For a duplicate key (legal JSON), this immediately overwrites the already-stored member with a discarded placeholder before the new value is even parsed. If the callback then rejects the new value, remove_discarded_value() finds the (still) discarded slot and erases the member entirely — losing the original value, not just skipping the new one.

parser_callbacks.md documents that "the parser will behave as if the discarded value was never read", which for a duplicate key should mean the first value stays in place:

auto drop_two = [](int, json::parse_event_t ev, json& v) {
    return !(ev == json::parse_event_t::value && v == 2);
};
json j = json::parse(R"({"a":1,"a":2})", drop_two);
// before: j.dump() == "{}"
// after:  j.dump() == "{\"a\":1}"

The same problem occurs when the rejected duplicate value is a container, rejected at object_end/array_end.

Fix

Add a small stash (duplicate_key_stash, a vector of slot-pointer/previous-value pairs) that key() populates when it detects the key already exists. The stash entry is dropped (superseded, no restore) once the new value is definitively accepted — for a scalar, in handle_value(); for a container, in end_object()/end_array()'s "kept" branch. It's consumed with a restore instead of an erase wherever a rejected value's placeholder would otherwise be erased — remove_discarded_value() (scalar rejection) and the discard branches of end_object()/end_array() (container rejection). Brand-new (non-duplicate) keys are completely unaffected, since no stash entry ever exists for them — that path is byte-for-byte unchanged.

Note: an separate, currently open PR (#5457) also touches remove_discarded_value() in this same file (an unrelated performance optimization, converting a linear scan to an erase-by-key lookup). Whichever of the two lands second will likely need a small rebase against the other.

Test plan

• Added tests/src/unit-regression2.cpp coverage for: both reproducers above (scalar and object), an array variant, accepted-duplicate regression guards (last value still wins, scalar and container), a brand-new-key-rejected regression guard, a two-level-nested duplicate key, and a three-occurrence key case.
• Ran the existing tests/src/unit-class_parser.cpp suite offline (10299 assertions, unaffected).
• Compiled and ran the two parser_callbacks.md doc examples (reject_duplicate_keys.cpp, parse__string__parser_callback_t.cpp) against the patched header — output byte-identical to the committed expected output.
• Regenerated single_include/nlohmann/json.hpp via make amalgamate.

Breaking change?

No breaking changes to the public API. This only changes behavior for the specific, previously-buggy case of a duplicate object key whose second occurrence's value is rejected by a parser callback — the member now correctly retains its first value instead of being erased, matching the documented "as if never read" contract. All other parsing/callback behavior (including the well-known "duplicate key keeps last value" rule when the callback accepts the value) is unchanged.

🤖 Generated with Claude Code