HardBreakSpaces normalization can emit a marker the parser will not honor (checkbox-only line swallows it)

[jbeda]

Emission-side sibling of #39. That fix made detection consult goldmark's per-line verdict; marker emission still assumes the configured spelling works in every context.

Repro (raw core; production is shielded by the render backstop)

$ printf '* [X] <br>\n0\n' | mdreflow --hard-breaks=spaces -
* [X] <br>
0

The unchanged output is the render backstop absorbing the failure. What the raw core does: pass 1 detects the real <br> break, normalizes it to the configured spaces spelling, and emits "* [X] \n0" — but the task-list extension's taskListRegexp (^\[([\sxX])\]\s*) swallows the checkbox, the trailing spaces, and the line ending, so the emitted marker renders as a soft break: the hard break the source had is lost. Pass 2's detection (now AST-confirmed, #39) correctly reads the emitted line as soft and joins — single-pass idempotency and render preservation both violated in the raw core.

Only the spaces spelling is affected: a backslash marker survives (\s* stops at \, leaving "\\\n" to parse as a break with an empty flagged text node), and <br> is raw HTML in any context. Only reachable with --hard-breaks=spaces and a hard break whose cluster prose reduces to exactly the checkbox shape on the paragraph's first line of a task-list item.

Why no soak ever finds it

deriveOptions in fuzz_test.go derives only Mode and MaxWidth from the input hash; HardBreaks is always the zero value (HardBreakBr). The spaces and backslash styles are never fuzzed at all — a coverage gap worth noting alongside #24/#26 regardless of this bug.

Fix directions

1. Per-position fallback, mirroring the existing first-output-line <br>→backslash fallback in writeParagraph: when attaching a spaces marker to a paragraph-first output line whose prose is exactly a checkbox shape and the paragraph is a list item's first child (p.Node parent chain answers this), fall back to the backslash spelling — render-preserving, one narrow position, style deviation only where the configured style cannot work.
2. Broader doctrine: validate emitted markers the way detectHardBreak invents a hard break after a task-list checkbox (raw trailing-space rule vs goldmark AST) #39 validates detected ones (reparse-based), which would also catch any future extension with the same swallowing behavior — likely overkill until a second context shows up.