Here's what I've been thinking of... It has two main differences: it encodes the type as a delta, which enforces a canonical encoding (lowest types go first), and it's slightly more compressed.

I think what you've currently got encodes:

• type = [0, 2**14), length(value) = 1,2,3 -- 2 bytes + length(value)
• type = [0, 2**14), length(value) = 0 or [4,253] -- 3 bytes + length(value)

I guess I find the "enforce canonical ordering by encoding as types as a delta" the most compelling part.

I found the RecordLowerBitMask stuff hard to follow as it's currently explained, fwiw.

...

Alternatively, perhaps doing a custom encoding with VarInt would be both clear and fine:

x = read_bytes(1)
y = read_VarInt()
type = last_type + (y << 6) + (x >> 2)
size = x & 0x3
if size == 3: size += read_VarInt()
value = read_bytes(size)

So your encoding is always:

• base_type_size : 1 byte (6 bits for type, 2 bits for size)
• extended_type : VarInt
• extended_size : optional VarInt (type_size & 3 == 3)
• value : size

That's still pretty compressed, and gives a canonical encoding/ordering...