fix: break over-long words in one pass instead of re-queuing the tail

[BrianWillows]

Summary

When a word is longer than the wrap width, the loop splits off one line-sized piece
and pushes the remainder back onto the queue:

case lineLength < wordLength: {
  const splitIndex = breakword(word, lineLength)
  const splitWord = [...word]
  words.unshift(splitWord.slice(0, splitIndex + 1).join(''))
  words.splice(1, 0, splitWord.slice(splitIndex + 1).join(''))
  break
}

Every pass then spreads the whole remainder ([...word]), measures it
(wcwidth(word) at the top of the loop), and slices/joins it again — so emitting
~width characters costs O(remaining). For a word of n characters that is
O(n²).

A single long token in otherwise ordinary text is enough to trigger it — a URL, a
hash, a base64 blob:

chars	size	before	after
4,000	4 KB	16 ms	2 ms
8,000	8 KB	43 ms	2 ms
32,000	31 KB	1,267 ms	5 ms
64,000	63 KB	5,958 ms	20 ms
128,000	125 KB	24,637 ms	14 ms

Narrower widths mean more iterations, so an ordinary 80-column terminal is in the
bad part of the curve: a 60 KB URL embedded in normal prose took ~1.4 s at
width: 80.

Fix

Break the over-long word into all of its chunks in a single pass, rather than
re-queuing the tail and re-scanning it each time. The chunk boundaries are computed
the same way (accumulate wcwidth until the next character would exceed
lineLength), so the output is unchanged.

words = chunks.concat(words) rather than unshift(...chunks) — a long word can
produce more chunks than the argument limit allows to be spread.

This removes the last use of breakword, so the now-unused import is dropped. Happy
to leave the dependency in package.json or remove it there too, whichever you
prefer.

Verification

• Existing test suite passes: 25/25.
• Output is byte-for-byte identical to the current implementation across 2,405
  generated inputs — CJK wide characters, tabs, mixed alphabets, widths 1/5/20/80,
  and explicit long-word cases ('a'.repeat(500) at width 20, '日'.repeat(60) at
  width 8, single characters wider than the line, etc.).
• The single-character-too-wide path still falls through to the existing
  config.errorChar handling.
• npm run lint output is unchanged apart from being one line shorter (the
  pre-existing no-fallthrough error at the deliberate fall-through remains).

Notes

Found and fixed with AI assistance (Claude). Happy to add a regression test that
asserts a long word wraps in bounded time if you'd like one.

[BrianWillows]

Nudge on this one - still merges cleanly against master as of today, CI green.

It replaces the re-queue-the-tail loop with a single pass, so wrapping one long token stops being quadratic. Existing tests pass unchanged.

Small enough to read in a minute; glad to adjust the style to match the rest of the file if it's off.