A hardbreaks block loses a rendered break when a lone + is fused onto the next word

What happens

In a :hardbreaks: document, every source newline in a paragraph renders as <br> (whitespace row A5). A lone + word must not end an output line (it would read as a hard break), so isDangerousAtLineEnd fuses the next word onto it. That fusion overrides a run the record held as a newline, and the rendered <br> is lost:

:hardbreaks:

* a
 +
  y

renders a<br> +<br> y; the formatter writes * a\n + y, which renders a<br> + y.

How it was found

The #304 lane measured a 1,152-case population (item text x {indented text, lone + at three columns, a trailing + line} x with/without :hardbreaks:). Main fails 344 of them (153 are this mechanism, all fidelity); the #304 change fails 136: 226 leave, 18 arrive, and every arrival is a :hardbreaks: document. The mechanism predates #304; #304 widens its reach from a + on the item's first continuation line to a + anywhere in the item's text.

Mechanism

The correct output needs the + to open AND close a line of its own, which needs both of its runs at once plus the item's continuation column, which wordsToAtoms does not have.

Candidate fix (a subtraction)

A hardbreaks block may never be re-laid-out at all, so make it a replayed block (a whole-block blockWhitespace row) instead of binding each run. That would delete row A5.

Ruling to take with it

Whether a lone + line inside a hardbreaks list item is a shape an author types. If not, this closes as contrived with the witness kept.

Related

#304, #301, #300.

Reproduce the hardbreaks witness, then trace how isDangerousAtLineEnd and wordsToAtoms handle the lone + and newline runs. Decide whether the hardbreaks block should be replayed or the shape should be ruled contrived; done requires a maintainer decision and a regression check for the chosen outcome.