Reject MessagePack/BSON binary subtypes that don't fit their wire format

[nlohmann]

Summary

byte_container_with_subtype::subtype() is a std::uint64_t, but the MessagePack ext-type byte and the BSON binary subtype byte are each a single byte on the wire. Both writers (include/nlohmann/detail/output/binary_writer.hpp) cast the subtype down to a single byte without any range check, silently wrapping modulo 256 instead of erroring — a silent data-corruption bug:

json::from_msgpack(json::to_msgpack(json::binary({1, 2}, 256))).get_binary().subtype();    // 0, not 256
json::from_msgpack(json::to_msgpack(json::binary({1, 2}, 70000))).get_binary().subtype();  // 112, not 70000

The valid range for both formats is actually the full 0–255 (a full unsigned byte), not just 0–127 — the MessagePack reader already round-trips 128–255 correctly via a signed/unsigned reinterpretation of the ext-type byte, and the BSON reader reads the subtype directly as uint8_t. Only subtypes ≥ 256 are affected. The CBOR writer already handles the full 64-bit range correctly (choosing 1/2/4/8-byte tag encodings), and the BSON writer already throws out_of_range.407-adjacent errors for other values it can't represent, so a check here is consistent with existing behavior.

Fix

Add a range check (subtype() > 255) before each of the two truncating casts, throwing a new out_of_range.413 (the next free out_of_range code after the existing .412) instead of silently truncating. A binary value with no subtype at all is unaffected.

Test plan

• Added tests/src/unit-regression2.cpp coverage: MessagePack/BSON subtype 0/200/255 still round-trip correctly (regression guard for the now-confirmed-correct 128–255 range); subtype 256/70000/300 now throw out_of_range.413 with the exact expected message; CBOR (already correct) remains unaffected at subtype 70000; a binary value with no subtype at all is unaffected.
• Documented the new exception code in docs/mkdocs/docs/home/exceptions.md (json.exception.out_of_range.413), matching the existing style.
• Regenerated single_include/nlohmann/json.hpp via make amalgamate.

Breaking change?

No breaking changes to the public API — this adds a new exception code (out_of_range.413) and only affects inputs that previously silently corrupted data (a subtype ≥ 256 being written and read back as a different, wrong value). No valid/previously-correct round-trip changes behavior; serializing a binary value whose subtype is ≥ 256 now throws instead of silently producing a wrong result.

🤖 Generated with Claude Code