VByte is a weird baseline to compare to: it is a byte-aligned encoding scheme, so it trades off some space efficiency for speed of decoding. The proposed method is bit-aligned, so it should be compared against bit-aligned encoding schemes.

For large integers, it is hard to beat Elias Delta coding [1], which asymptotically uses log(n) + o(log(n)) bits, and in practice it breaks even with most other encoding schemes quite early on.

More importantly, Elias Gamma and Delta are complete, meaning that there is no redundancy (another way to look at it is that any sequence over the alphabet is decodable). VByte is complete over the byte alphabet. Any complete code is optimal for some integer distribution (see "implied probability" in the Wikipedia page).

So if your integer distribution is heavier on the large integers, there are already plenty of complete encodings to pick from, and almost certainly one that fits well the distribution.

The scheme proposed here, instead, is not complete, as mentioned in the README ("this method does introduce some redundancy, particularly when sequences must be padded to prevent unintentional "101" patterns"), so it cannot be optimal for any distribution.

The Wikipedia page on the Kraft–McMillan inequality [2] explains this in more detail.

[1] https://en.wikipedia.org/wiki/Elias_delta_coding [2] https://en.wikipedia.org/wiki/Kraft%E2%80%93McMillan_inequal...