feat(dedupe): URL canonicalization for cross-source linking

[phjlljp]

Summary

Problem: Same article shared across sources isn't cross-linked

cross_source_link() detects when the same story is discussed on multiple platforms (e.g., Reddit + HN + Bluesky) and annotates items with [also on: ...] tags. Currently it relies entirely on text similarity (Jaccard on character trigrams and word tokens). This misses a common case: the same URL shared across platforms with different tracking parameters.

For example, these three URLs all point to the same article but would not be cross-linked today:

• Reddit: https://www.example.com/article?utm_source=reddit&ref=share
• HN: https://example.com/article
• Bluesky: https://www.example.com/article?fbclid=abc123

The titles may also differ (Reddit editorializes, HN uses the original, Bluesky truncates), so text similarity alone doesn't reliably catch these.

Solution: URL canonicalization as a fast-path check

This PR adds a canonicalize_url() function that normalizes URLs before comparison:

• Strips tracking/analytics query parameters (UTM, fbclid, gclid, msclkid, igshid, etc.)
• Removes www. prefix and lowercases the host
• Strips trailing slashes and URL fragments
• Rejects non-HTTP schemes (file://, javascript:, etc.)
• Sorts remaining query parameters for consistent comparison

cross_source_link() now pre-computes canonical URLs for all items (O(N), using stdlib urlparse) and checks URL identity first. If two items from different sources have the same canonical URL, they're cross-linked immediately without computing text similarity. This is both faster (string equality vs. Jaccard) and more accurate for URL-shared content.

The existing text similarity path remains as a fallback for cases where the same story is discussed but linked differently (e.g., a Reddit self-post discussing an HN thread).

Design decisions

• Conservative tracking param list: Only strips params that are unambiguously tracking-related. Generic params like source, context, s, t are intentionally kept because they serve functional purposes on many sites (e.g., Reddit's ?context=3 for comment threading).
• Pre-computed, not inline: Canonical URLs are computed once per item before the O(N²) pairwise loop, not inside it.
• Scheme restriction: Non-HTTP URLs return None and are excluded from URL matching (but still participate in text similarity).

Test plan

• Existing test suite passes (461/465, 4 pre-existing)
• URLs with UTM/fbclid params canonicalize to the same string as bare URLs
• ?context=3 (Reddit) and ?source=rss (functional) are preserved after canonicalization
• file://, javascript://, and ftp:// URLs return None
• Cross-source linking correctly links Reddit, HN, and Bluesky items sharing the same article URL with different tracking params (verified bidirectional cross-refs)

🤖 Generated with Claude Code

[mvanhorn]

See my reply on #76 covering all three of your PRs. Can't commit to anything right now with the v3.0 refactor underway, but I'll consider each one once it lands.