Handles reasoning model output with analysis text.

One site crashing doesn't stop others.

Run merger against a temp directory with sample yaml/txt files.

Each keyword should belong to exactly one region to avoid ambiguity.

Ensure src.llm was deleted and not referenced.

When pattern fails >=5 times, link_pattern gets reset to None.

LLM generates valid pattern → saved to config.

Chinese date "6月18日" in title → sorted correctly.

URL path fallback when title has no date.

Relative href gets joined to base URL.

Semaphore cap = 2, max 2 concurrent.

Config.crawl.max_articles = 3 limits the result.

Retries 3 times, returns None on total failure.

link_pattern matches → returns links, no LLM.

No pattern → LLM extracts links.

All 3 sites processed, results collected.

Only the targeted site runs.

Category, pagination, and root links are excluded.