The problem is that this breaks down once you try to use SIMD instructions. I'd developed a similar kind of approach to encoding integers (and ieee774 floats) a couple of years ago (first byte encodes length and first bit of data: https://github.com/kstenerud/bonjson/blob/05b91f6fe7d6b07186... ). It was very clever and used compiler intrinsics to get the length in 1 instruction, so 2 instructions got you the final value, with no branches.
But testing proved that when you move to SIMD instructions, ULEB128 (https://github.com/kstenerud/bonjson/blob/main/bonjson.md#ty...) or sentinel values (https://github.com/kstenerud/bonjson/blob/main/bonjson.md#lo...) win every time because of the parallelization opportunities.
The true irony is that even SIMD text parsing would outperform this! SIMD is that powerful.