fix(jira): preserve escaped emphasis delimiters in jira_to_markdown

[barrygfox]

Problem

JiraPreprocessor.jira_to_markdown() pairs emphasis delimiters positionally with ([*_])(.*?)\1, ignoring wiki escape markers. Stored content containing escaped intraword underscores is rewritten on every read-back:

• stored wiki bytes: QUALITY\_GATES\_LLM\_ENABLED
• returned "markdown": QUALITY\*GATES\*LLM\_ENABLED

Each round-trip through any read tool rewrites \_ to \*, silently corrupting content for MCP clients and any agent that treats the response as ground truth. Tracked by #1610.

Solution

Delimiters preceded by a backslash are wiki escapes written by this module's own markdown_to_jira (intraword \_ protection), not markup. The pairing regex now rejects escaped delimiters via negative lookbehind, so they can neither open nor close a span:

r"(?<!\\)([*_])(.*?)(?<!\\)\1"

Changes

src/mcp_atlassian/preprocessing/jira.py

Change	Why
Emphasis regex gains (?<!\\) lookbehinds on open/close delimiter	Escaped delimiters are literal text, not markup

tests/unit/preprocessing/test_preprocessing.py

• test_jira_to_markdown additions — escaped \_ / \* sequences survive conversion unchanged; unescaped emphasis still converts

What Does Not Change

• Unescaped emphasis conversion is byte-for-byte identical (*bold* → **bold**, _italic_ → *italic*)
• Write path (markdown_to_jira) untouched
• No new dependencies; no API/config changes

Breaking Changes

None.

Regression Tests

Fails before the fix (tests/unit/preprocessing/test_preprocessing.py::test_jira_to_markdown, against 436b37e):

E       AssertionError: assert 'QUALITY\\*GA...LLM\\_ENABLED' == 'QUALITY\\_GA...LLM\\_ENABLED'
E         - QUALITY\_GATES\_LLM\_ENABLED
E         + QUALITY\*GATES\*LLM\_ENABLED

Result: FAIL before, PASS after this commit.

Test Plan

• pytest tests/unit/preprocessing/ -q — 173 passed
• pytest tests/unit -q — full suite green on the combined stack

Fixes #1610

[mcp-atlassian-maintainer]

Merged, thanks @barrygfox. The focused regex change and regression cases preserve escaped Jira identifiers while keeping real emphasis conversion intact.