…5-sonnet-thinking-451d

…n categorized commits

Previously, categorized commits would not check the commit message body
for the #body-in-changelog magic word, only the PR body. This was
inconsistent with uncategorized commits (leftovers) which correctly
checked both. Now both code paths behave consistently.

- Test that PR body is NOT included without magic word (both categorized and uncategorized)
- Test that PR body IS included with magic word
- Test that commit body is used as fallback for magic word in both categorized and uncategorized commits
- Test that PR body takes precedence over commit body when both have magic word
- Test consistency between categorized and uncategorized commits

…kdown

When a Changelog Entry section contains plain text (no bullets), multi-line
content is now joined with spaces to produce valid markdown output.

Previously, newlines were preserved which resulted in broken output like:
- This is a multi-line
changelog entry that
spans several lines. by @user in [#1](url)

Now produces:
- This is a multi-line changelog entry that spans several lines. by @user in [#1](url)

Also adds regression test for this case.

PR descriptions from GitHub may have Windows line endings (CRLF).
The regex patterns using $ to match end-of-line don't match before
\r characters, causing bullet points to not be recognized as
separate entries.

This fix normalizes all line endings to LF before parsing.

Previously, nested bullets like:
  - Add feature
      - Sub-item 1
      - Sub-item 2

Would render as:
  - Add feature by @user in [#1]
    Sub-item 1
    Sub-item 2

Now correctly renders as:
  - Add feature by @user in [#1]
    - Sub-item 1
    - Sub-item 2

Nested items do NOT get author/PR attribution - only the top-level entry does.

- Use marked library for proper markdown parsing instead of complex regex
- Correctly handles CRLF line endings, nested lists, and edge cases
- No dependencies (marked has zero dependencies, ~430KB unpacked)
- Added Jest moduleNameMapper to use UMD build for compatibility
- Updated tests to reflect correct markdown behavior (indented bullets
  after a paragraph without blank line are paragraph text, not a list)

- Convert locateChangeset, extractChangeset, removeChangeset, and
  prependChangeset to use marked lexer instead of complex regex
- Extract common test fixtures to src/utils/__tests__/fixtures/changelog.ts
- Remove duplicate TestCommit interface
- Simplify test markdown to use proper non-indented strings
- Reduce test file size by ~100 lines
- Update jest config to exclude fixtures directory from test discovery

BREAKING: Removed changelog.test.ts (4237 lines) and replaced with:
- changelog-file-ops.test.ts (210 lines) - findChangeset, removeChangeset, prependChangeset
- changelog-generate.test.ts (535 lines) - generateChangesetFromGit with snapshot testing
- changelog-extract.test.ts (195 lines) - extractScope, formatScopeTitle, extractChangelogEntry
- changelog-utils.test.ts (144 lines) - shouldExcludePR, shouldSkipCurrentPR, getBumpTypeForPR

Total reduction: 4237 → 1444 lines (~66% reduction)

The snapshot tests capture the full output format without inline expected strings,
making the tests much more maintainable and readable.

…5-sonnet-thinking-451d