fix: unify export content across formats (title, soft breaks, comments)

[rogerdigital]

Problem

Exported content diverged from the source in ways that differed across the four formats. Investigation (filename vs content) found the filenames are already consistent (all formats write the same outputFilePath from ExportRunner). The issues are all in content layer:

Fixes

1. Title duplication with body's first H1

When a document started with a level-1 heading, every format prepended its own title (# / <h1> / Title style) on top of that heading still present in the body → duplicate title.

Fix: DocumentAssembler now extracts a leading # Heading as the document title and removes it from the body (extractLeadingH1). When there's no leading H1, the filename-derived title is used unchanged. Affects all formats uniformly.

Before	After
# filename + # original heading (duplicate)	# original heading (no duplicate)
# filename (no H1 in body)	# filename (unchanged)

2. Markdown embed empty alt text

LinkRewriter.formatEmbed emitted ![](path) with empty alt for the markdown-bundle profile. Obsidian (and many viewers) don't render images with empty alt, so images were invisible.

Fix: Now uses the link text as alt (![image.png](path)), matching the docx/html/pdf branches.

3. docx soft line breaks became separate paragraphs

parseMarkdownToParagraphs split on every \n, so adjacent non-blank lines (Obsidian soft breaks within one paragraph) each became their own <w:p> with large paragraph spacing.

Fix: Consecutive plain-text lines are now aggregated into one paragraph, separated by <w:br/>, matching how Obsidian renders them.

4. docx rendered source-path comments as literal text

<!-- source: path --> lines appeared as visible body text in docx.

Fix: HTML comment lines are skipped during paragraph parsing.

Not changed

• Image size (![[img.png|258]]): the |width Obsidian syntax is stripped upstream in LinkRewriter for all formats. Standard Markdown cannot express it; preserving it would require HTML <img width> in markdown-bundle, which is out of scope.

Tests

164 pass (added extractLeadingH1, soft-break aggregation, comment-skipping coverage).