don't buffer commonmark output unless necessary.

[kivikakk]

The status quo is that cm will buffer all output in self.output, which gets written out all-at-once to the provided fmt::Write.

If we're not wrapping (options.render.width > 0), and we're not using the experimental escape minimizer (GH-523), we don't need to buffer output at all.

If we're wrapping, we need to buffer at most until a line break or wrapping event occurs.

If we're using the experimental escape minimizer, we still have to buffer everything.

It was annoying to move from doing things this way (made easy by the reference C implementation) because we need to know some properties of the last few bytes written out; whether up to the last two bytes were linefeeds, and whether the last byte was an ASCII digit. We keep a window of the last two bytes written out to furnish this.

[github-actions]

Command	Mean [ms]	Min [ms]	Max [ms]	Relative
./bench.sh ./comrak-3bdc700	104.4 ± 1.3	102.1	107.7	1.76 ± 0.06
./bench.sh ./comrak-main	103.9 ± 1.1	101.4	106.5	1.76 ± 0.06
./bench.sh ./pulldown-cmark	59.2 ± 1.9	57.8	70.7	1.00
./bench.sh ./cmark-gfm	92.0 ± 6.9	81.8	104.2	1.55 ± 0.13
./bench.sh ./markdown-it	260.6 ± 3.0	253.9	268.2	4.40 ± 0.15

Run on Sat Nov 8 12:48:21 UTC 2025