A '#' that begins a comment must be preceded by whitespace or open the
line; glued to earlier same-line content (e.g. "value"# c, [ a, b ]#c,
>#c, a flow entry then ,#c) it is neither a comment nor, in those
positions, plain-scalar content, so the document is malformed. Such input
was previously accepted as a comment.
A '#' that opens a line, follows a space, or sits inside a plain scalar
(a#b, where it stays literal content and never reaches this check) is
unaffected. The scanner tracks whether a real token has been scanned on
the current physical line, reset on each line break, rather than comparing
token line numbers: the streaming tokenizer drains its token buffer
between scans, and a block scalar's blank line emits an empty scalar
attributed to the next line, either of which would misread as content.
Measured against the official YAML test-suite, parse-level conformance
rises from 243/333 (73.0%) to 247/333 (74.2%): four wrongly-accepted
invalid documents are now rejected, with no valid document regressing.
The scalar-scanning routines move to Scanner.Scalars.swift to keep the
file within the length limit.
The scanner dropped every blank line within and after a block scalar, so
a literal block lost internal blank lines, a folded block turned a blank
line into a space rather than a newline, and the keep indicator could not
preserve trailing blank lines. Capture blank lines as empty content in
both the batch scan loop and the streaming TokenCursor loop, which had
drifted as duplicate copies. Rework blockScalarValue to apply literal
versus folded line semantics and clip/strip/keep chomping over the
captured lines; content without blank lines is unchanged. Re-pin the
multiline-scalar test whose expectation encoded the old dropped-blank
behavior, and add a chomping-and-folding test matrix.
Add structured validation diagnostics, JSON and Yams comparison gates, and lazy token/event parsing so production behavior is measured without retaining full intermediate parser arrays.