fix(jira): stop prose plus signs corrupting markdown read-back

[barrygfox]

Problem

Unpaired plus signs in Jira prose, especially version numbers such as Python 3.13+ and urllib3 2.6+, could be paired across text by the inserted-text rule. That emitted an <ins> tag, triggered whole-document HTML conversion, escaped existing Markdown characters, and removed the plus signs from read responses.

The HTML guard also checked content only after jira_to_markdown(). Converter-generated tags therefore still looked like source HTML, so the guard did not prevent the second conversion.

Solution

1. Inserted-text delimiters adjacent to word characters are treated as prose: (?<!\w)\+([^+\n]+)\+(?!\w).
2. clean_jira_text() records whether the original Jira source contains HTML outside protected code spans before Jira markup conversion.
3. Whole-document HTML conversion runs only when HTML came from that original source. Tags emitted by Jira conversion remain valid inline HTML in the Markdown result.

Changes

• src/mcp_atlassian/preprocessing/jira.py
  • Preserve unpaired version-number plus signs.
  • Detect source HTML without mistaking Jira or Markdown code spans for HTML.
  • Keep converter-generated <ins>, <cite>, <sup>, and <sub> tags from triggering a second conversion.
• tests/unit/preprocessing/test_preprocessing.py
  • Cover stray plus signs alongside Markdown emphasis.
  • Cover literal plus signs in smart-link URLs.
  • Prove generated HTML does not trigger conversion, including when literal HTML appears inside a code span.
  • Preserve genuine inserted-text markup.

Verification

• uv run pytest tests/unit/preprocessing/ -q: 175 passed
• uv run pytest tests/unit/ -q: 3842 passed, 5 repository-standard skips
• uv run pytest 'tests/e2e/test_jira_auth_matrix.py::TestJiraReadOperations::test_get_issue[basic]' --dc-e2e -vv: 1 passed, no skips
• uv run python scripts/generate_tool_docs.py --check: passed
• Ruff check and format check on changed Jira/test files: passed
• mypy hook on changed Jira/test files: passed

A broader DC run reached 102 passed, 1 failed, and 1 skipped. The failure is an unrelated JSM fixture problem: the pre-seeded JSMDC-2 issue returns 403 from the service-desk comment endpoint.

Related: #1610