Round 3 fixed a phantom-space idempotency bug by dropping any boundary
whitespace containing a newline, but that also stripped ordinary
hand-wrapped text like "hello\n<strong>", merging words on a single pass --
directly contradicting this formatter's own "does not reflow text"
contract. The correct discriminator is what the whitespace borders, not
whether it contains a newline: a run between two inline-level things (text,
<strong>, <a>, ...) is always significant and must survive regardless of
newlines, while a run touching a block-tag boundary carries no rendered
meaning and is always dropped. Since this formatter's own emitted
indentation is only ever inserted next to a block tag, that rule also
resolves the original phantom-space bug without any newline special-casing.
tokenize() now peeks each upcoming tag's name once (reused for both the
preceding text run's decision and the tag's own processing) so pushTextToken
can see what's on both sides of a whitespace run.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
shadowdaoandClaude Opus 5
45ee004672