fix(i18n): keep a rejection sample from breaking the catalog it annotates

record() quotes the rejected translation into a PO comment and guarded the known hazard, a trailing backslash, with rstrip. The guard sits too early: babel wraps the finished comment across several # lines afterwards, so a backslash inside the string can land at a break it never saw. In docs/tr the wrap fell inside an escaped \\\\ of the source, gettext read the dangling backslash as a line continuation, swallowed the msgid below it, and msgfmt reported 56451:7 syntax error on an msgstr with no counterpart. docs/zh failed the same way.

Removing every backslash instead of the last one makes the hazard impossible wherever the wrap falls, and a newline becomes a space rather than an escape that reintroduces one. A rejection sample quotes what the engine said; it needs no escapes.

Reproduced and closed: both catalogs failed msgfmt before and pass after. A sweep over the tree finds no syntax error in any of the 180+ catalogs, though 24 still carry the same dangling backslash where the next line happens to be another comment.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

utils/i18n/translate.py

@@ -230,7 +230,7 @@ def record(message: Message, rejected: Rejected | None) -> None:
     )
     message.user_comments.append(URL_SUPPRESSION)
     message.user_comments.append(
-        f"{REJECTED_PREFIX} {rejected.text}".replace("\n", "\\n").rstrip("\\")
+        f"{REJECTED_PREFIX} {rejected.text}".replace("\\", "").replace("\n", " ")
     )