[mananaysiempre]

  static int64_t read_sprefix(unsigned char **buf) {
    uint64_t x = *(uint64_t *)*buf;
    unsigned n = stdc_trailing_zeros(x) + 1;
    assert(n <= 8); /* handles values up to 2**56 - 1 */
    *buf += n;
    return (int64_t)(x << 64 - 8*n) >> 64 - 7*n;
  }

Seems pretty OK too and doesn’t force a branch per byte. (Imagine a memcpy instead of the first line if the strict-aliasing UB annoys you.) I guess it does do somewhat more work if the majority of the inputs fits in a byte.