Techniques

The field walk is a compile-time loop over reflect[T].field_count(), field_types(), and field_ref. Scalars are little-endian. On a little-endian host each multi-byte scalar is one unaligned store. A list of Int32, Float64, or another fixed-width number is one memcpy on this little-endian host. Strings are a u32 length and then the UTF-8 bytes.

A struct whose fields are only scalars and strings is measured once. encode reserves that many bytes, then writes them through one pointer. On a little-endian target each 2-, 4-, and 8-byte value is one unaligned Pointer.unsafe_store. read_string copies into String(unsafe_uninit_length=) with memcpy. An all-ASCII string is accepted by OR-ing its bytes, eight at a time on a little-endian host, and only a byte past 0x7F runs the multi-byte UTF-8 checker. A struct that contains a list, an optional, or another struct still reserves inside those values, so a short nested field cannot write past the bytes it owns. After the first call the buffer is large enough, and later calls do not allocate.

decode is the safe entry point. It rejects a length past the end of the buffer before it allocates. Reader.u32_fast and str_fast skip those checks. The batch measurement uses them, and they are only safe for a buffer whose bounds are already known.

Dict keys are sorted by their UTF-8 bytes before they are written. Iteration order of Dict is not part of the wire format.

A custom type can conform to KilnCodec and replace the derived field walk. Scalars, String, List, Optional, and Dict keep the layout in the spec.