Techniques

Frozen values are positional. The schema stays in the program. Dict keys are sorted by their UTF-8 bytes so two encodes of the same mapping match. Floats keep every bit, including the sign of zero and the payload of a NaN.

The tensor directory is one Frozen value. Its magic is KLNT and its version byte is 2. Descriptors are fixed width. Payloads start on a 64-byte boundary measured from the start of the directory. Decode borrows those bytes from the caller's buffer. Tensor.data copies them out when the caller asks for an owned array.

A count is checked against the unread bytes before Kiln allocates storage for the elements. The default cap is 16,777,216 counts and 16 MiB for one string or byte run. Pass cap: to raise it. Nesting deeper than 64 containers is LIMIT.

A frame is sixteen bytes: KLN0, four zero bytes, and the FNV-1a 64 hash of the canonical type text, then one Frozen value. A hash that does not match the schema the caller passed is SCHEMA. typeHash computes that hash. The token for a tensor field is the ASCII bytes tensors.

Writer.encode resets the length and keeps the existing storage. Call it in a loop when the same process encodes many records. Numeric list bodies and tensor payloads are copied in bulk. Field names on a struct are not reallocated; the schema already owns those strings.

bash scripts/run-microbench.sh prints nanoseconds for a small struct batch and for one 1024-element f32 tensor, encode and decode.