Reuse a scratch buffer when reading child output

[tas50]

Description

read_nonblock(READ_SIZE) allocates a fresh 4 KB String on every read. For a command with a lot of output that's nearly all of the garbage ShellOut produces — an 8 MB stdout means ~2000 needless allocations that exist only to be copied into @stdout and thrown away.

IO#read_nonblock takes an optional buffer to read into, so this hands it one and reuses it across reads.

That's only safe because #<< copies the bytes into @stdout/@stderr. A live stream is different — it's documented as anything responding to <<, and it's free to hold onto whatever it's handed. Reusing one buffer there would silently alias every chunk together (an Array used as a live_stream would end up with N references to the same String). So when a live stream is attached, each read still gets its own String, exactly as today.

That split is what the new #drain helper expresses, and it has the side benefit of collapsing the three near-identical read loops into one.

Benchmark

cat of an 8 MB file, 25 runs, ruby 4.0.6, arm64-darwin. objects/run is GC.stat(:total_allocated_objects) across a single run_command.

	median	min	objects/run
before	30.76 ms	16.81 ms	1997
after	16.10 ms	13.42 ms	100

48% faster with 95% fewer allocations. The min column (13.42 ms vs 16.81 ms) is the GC-free floor — the rest of the median gap is GC pressure the allocations were causing.

Full matrix, same machine:

case	before (median / objects)	after (median / objects)
tiny stdout (echo hi)	6.94 ms / 51	6.23 ms / 50
1 MB stdout	9.52 ms / 288	8.39 ms / 44
8 MB stdout	30.76 ms / 1997	16.10 ms / 100
8 MB stdout + live_stream	30.38 ms / 3957	31.64 ms / 3960
1 MB stderr	22.82 ms / 311	23.17 ms / 48

The live_stream row is deliberately flat — that path runs the same code it does today.

Verification

Output is byte-identical, checked against main rather than assumed:

• 8 MB of stdout — same SHA256, same size, same encoding (UTF-8)
• interleaved binary + multibyte stdout and stderr ("\xff\xfe café" / "err\xc3\xa9") — same SHA256, same encoding (ASCII-8BIT) on both streams
• live_stream receiving 1954 chunks — same joined SHA256, and explicitly asserted the chunks are not aliased to each other
• pre-exec failure still propagates (Errno::ENOENT), which exercises the process_status pipe's buffer

Plus:

• bundle exec rspec — 147 examples, 0 failures (unchanged from main)
• bundle exec cookstyle --chefstyle -c .rubocop.yml — no offenses

Note

The first commit is a spellcheck fix. The spellcheck job only scans files a PR touches, so any change to unix.rb fails on identifiers that have been in the file for years (pgid, sgids, WNOHANG, endgrent, …). It also flagged proccess, which is a genuine typo in a comment, so that's fixed rather than added to the dictionary. The same commit is in #282 — whichever lands first, the other needs a trivial rebase.