Add resources/excluded_phrases.txt (skip links, privacy/consent, legal
boilerplate) and match it case- and punctuation-insensitively as a
contiguous run of words. Phrases live in a data file so they can be
extended without touching code.

News sites often reject minimal-bot requests; send a full browser header
set (User-Agent, Accept, Sec-Fetch-*, client hints). Accept-Encoding is
left to requests/urllib3 to avoid undecodable brotli responses.

- Track source domain and publication time on each headline
- Cluster near-duplicate headlines into stories (union-find)
- Output stories covered by >=N distinct sources (--min-sources)
- Add --since/--until date window
- Ignore __pycache__ build artifacts