Bounds-checked little-endian (and two big-endian) reads from a ByteString at an offset: what the value decoders use once a value's length has been checked against its type.

No Get, no intermediate ByteString, and no pointers here: each reader checks that the n bytes at off lie within the buffer (calling error if they don't, as index does: a read can never touch memory outside the buffer, whatever the caller got wrong), then assembles the value from its bytes. Raw pointer operations, for the character codecs, are in Hmssql.Tds.Bytes.Unsafe.

The parser and the scalar codecs read with these. A field conversion receives a whole value, so it never needs an offset into a packet.