fix(shopify): parse malformed llms.txt

[Jadenzzz]

🚥 Resolves ISSUE_ID

🧰 Changes

The llms.txt extractor was silently dropping the majority of links on many real docs sites. Importing https://shopify.dev/docs produced 2 off-origin "Deploy to Fly.io / Render" pages as if they were the entire Shopify site, instead of the ~hundreds of actual Shopify references the file contains.

Five layered fixes to parseLlmsTxt / fetchLlmsTxt, plus one thread-through in the prompt builder.

Fixes

• Accept * and + bullets, not just - since CommonMark allows all three; the llms.txt spec doesn't pin one. Shopify uses *; under the old regex 16/18 of its bullets were silently dropped.
• Recognize ##–#### as section starts — Shopify's file has only 2 ## headings, with 18 ### and 54 #### carrying the real structure. The old h2-only check collapsed everything into a single mega-section.
• Harvest inline [text](url) links from paragraph prose — Some files (Shopify, again) keep most refs in narrative text rather than bullet lists. Added a second sweep that runs on non-bullet, non-code-fence lines. Items dedupe by URL across the parse.
• Code-fence-aware-> won't extract URLs from inside ``` blocks.
• Same-origin filter (opt-in via sourceUrl) — When the source URL is known, all items must share its origin. Drops cross-references to GitHub, partner deploy guides, blog posts, etc. Strict origin (no subdomain coalescing).
• Strip matched markdown emphasis from display text - **Extension targets**, *italic*, __bold__, and combined forms like **_x_** are unwrapped. Asymmetric input (**foo) is left untouched.

🧬 QA & Testing

Provide as much information as you can on how to test what you've done.

Loom

https://www.loom.com/share/4bacf2d0195040f4a10bd2efaaf1dff4