deps: require trafilatura >=2.2 — the #882 canary fired, close the loop

[Wynelson94]

The canary worked

tests/test_content.py pinned trafilatura's broken nested-emphasis serializer on purpose, so an upgrade would fail loudly rather than silently change output. It fired:

• trafilatura 2.2.0 published 2026-07-31; adbar/trafilatura#882 — the issue this repo filed — closed the same day
• The fix is real, verified in xml.py source rather than release notes: _md_wrap() keeps flanking whitespace outside the emphasis marker (so the closing delimiter is right-flanking per CommonMark) plus _collapse_emphasis() for nested hi chains
• The pinned test failed on cue, and is now flipped to pin the corrected serialization as a permanent regression guard

CI was not red yet only because it last ran 07-24 and triggers on push/PR — it would have gone red on the next push regardless of this branch.

Adoption was measured, not assumed

Replayed all 194 cached pages under both versions (zero network — the cache stores raw HTML):

better under 2.2.0	77
worse	42
identical	75
corpus total	+2.89%
catastrophic (≥50% loss)	0

2.2.0 repairs three of the five silent-omission failures the stub guard was built on:

• sbir.gov/awards 421 → 3,155 chars — previously returned only the "official website of the United States government" banner with zero award data
• atlas.nomic.ai/pricing 259 → 1,661
• NOAA SBIR archive 361 → 548

The stub guard now fires on 2 of 194 (was 5 of 162) because the corpus got better, not because the guard got weaker. Calibration table updated in place; thresholds unchanged.

Two pins moved — deliberately, neither quietly

article_peripheral regenerated. It padded its body with 14 byte-identical <p> tags; 2.2.0 deduplicates identical blocks, so the body collapsed 1,973 → 166 chars and the page started tripping the stub guard — while the real page it models went the other way (1,827 → 13,560, +642%). The synthetic had stopped modelling reality in both magnitude and direction. scripts/gen_article_peripheral_fixture.py is now its reproducible source, and the rebuild re-verifies every clause of the original contract (4 figures, max 3 per 1,500-char window, silent under shipped guards, warns at floor=1).

Its tests were why nobody noticed: the body assertion checked only "Join Forces" — which the page title satisfies — so a 92% body collapse passed CI. It now pins real body sentences, total guard silence, and that the silence comes from the cluster floor rather than a toothless page.

quo's collapsed-column assertion narrowed. Quo is just as lossy (still 4 of 24 figures, 1 of 3 plan names) but 2.2.0 emits the repeated cell once instead of three times, so the repetition signature the guard keys on is gone — on quo and on every page in the replay. The dropped-price guard still catches quo, so the page is not silent. The substantive loss is pinned separately and the coverage gap is recorded in CHANGELOG under Known issues rather than papered over.

⚠️ Known issue this surfaces: the collapsed-column guard now has zero real-page coverage. Tracked for the follow-up guard-coverage pass, not fixed here.

Verification

• ruff · mypy · 252 tests · check_consistency.py — all green (the consistency gate caught the 247→252 count drift, as designed)
• Full live eval re-run on 2.2.0 per the standing post-upgrade protocol — report added to this PR

🤖 Generated with Claude Code