Fix to_bjdata() silently truncating out-of-range _ArrayData_ elements

[nlohmann]

Summary

The BJData ndarray writer (write_bjdata_ndarray in include/nlohmann/detail/output/binary_writer.hpp) validated the kind of each _ArrayData_ element (integer vs. float) against the declared _ArrayType_, but not its range. Each element is written with a narrowing cast, e.g. write_number(static_cast<std::uint8_t>(el.get<std::uint64_t>()), true) for _ArrayType_ = "uint8". An out-of-range element (e.g. 256) was silently wrapped by that cast instead of being reported — no error, no fallback.

json j;
j["_ArrayType_"] = "uint8";
j["_ArraySize_"] = json::array({2});
j["_ArrayData_"] = json::array({1, 256});   // 256 does not fit uint8
auto v = json::to_bjdata(j);
// json::from_bjdata(v).dump() used to give "[1,0]" instead of round-tripping

Fix

Each _ArrayData_ element is now range-checked against the C++ type named by _ArrayType_ before being cast:

• integer types (uint8/int8/uint16/int16/uint32/int32/uint64/int64, plus char/byte which share the uint8 range) are checked with the existing value_in_range_of<TargetType> helper, handling both signed and unsigned JSON number storage (so a negative value is correctly rejected for an unsigned target).
• single (32-bit float) is checked so that a finite double which would overflow to infinity when narrowed to float is rejected.
• double needs no additional check since it already spans the full range of the internal float representation.

When any element is out of range, the function falls back to the same plain-object encoding already used by this function for other invalid-annotation cases (mismatched kind, non-array _ArraySize_, overflowing dimensions, etc.), so the value round-trips through from_bjdata(to_bjdata(j)) instead of being silently corrupted.

Tests

Added a new section "ndarray with out-of-range _ArrayData_ elements stays as object" in tests/src/unit-bjdata.cpp covering:

• an out-of-range uint8 element (256) falls back and round-trips
• an out-of-range int8 element (200) falls back and round-trips
• a negative element under an unsigned type (uint16) falls back and round-trips
• a single (float) element that overflows to infinity (1e40) falls back and round-trips
• in-range boundary values (uint8 0/255, int8 -128/127, single 1.5) still use the compact ndarray encoding

Ran the full unit-bjdata suite offline (compiled against include/ with a stub test_data.hpp): 693935/693936 assertions pass; the one failure and the one skipped test case are pre-existing and unrelated (they require downloaded test data, unavailable in this offline setup).

Breaking change?

No breaking changes. This only affects the BJData ndarray fast-path encoding for annotated objects whose _ArrayData_ previously contained values outside the range implied by _ArrayType_; such objects are now encoded as plain objects (as they already are for the other invalid-annotation cases handled by this same function) instead of emitting silently corrupted data.

Fixes #5403.

— opened by Claude Code on behalf of @nlohmann