fix: treat only spaces and tabs as blank when picking the streaming parser's stable boundary

[Ryan VanderMeulen (rvandermeulen)]

Problem

The block parser treats a line as blank only when it holds nothing but spaces and tabs (MarkdownParserUtil.isEmptyOrSpaces), but lastBlankLineEndOrNull, which picks the streaming parser's stable boundary, tests isWhitespace(), which also accepts \r, a non-breaking space and other whitespace. With CRLF line endings, line one\r\nline two\r\n\r\nnext\r\n is a single paragraph to a full parse, since \r\n\r\n is not a blank line to the block parser. Streamed one character at a time, the boundary settles after \r\n\r\n and next becomes a second paragraph, so the streamed tree depends on how the text was chunked. A line holding only a non-breaking space does the same with LF line endings.

Approach

Use the block parser's definition of a blank line in lastBlankLineEndOrNull. The range also drops the two \n endpoints, which were only there because isWhitespace() accepts them.

With CRLF line endings the boundary now never settles on a blank line, so the whole document stays in the unstable tail and is reparsed on every append, as it would be for a document without blank lines. That is consistent with what a full parse produces; recognizing \r\n as a line ending in the block parser would restore incremental parsing but is a much larger change.

Tests

A CRLF document and an LF document with a line holding only a non-breaking space, each streamed one character at a time, produce the same top-level children as a full parse. Each input fails on its own without the fix.