#29 co-core 0.32.0: XML and feed extractors, refusals as goldens

[gregoryfoster]

Refs #29. (Refs, not Closes: Watcher's acknowledgement on watcher#366, the deploy and the live gate come after the merge.)

Plan: docs/plans/2026-10-12-29-co-core-0.32.0.md. It was reviewed before implementation, and every proposed case was run through 0.32.0's chain.

What changes

• The pin moves 0.19.7 → 0.32.0. This is the catch-up jump, carrying cannobserv#523. processor_version becomes 0.32.0+1; LOCAL_GENERATION stays 1 (ask 2). The lock moves only co-core, co-core-aio and co-core-sync: no dependency of theirs changes.
• No change to the code that runs, except a docstring. The four XML essences now dispatch to XmlExtractor. A spec co-core can't honour raises UnsupportedExtractionError. That error propagates, the child reports it raised, and the handler publishes extraction_error, terminal (spec §4).
• Refusals are goldens. gen_parity_goldens.py records a refused case as {"refused": "UnsupportedExtractionError"}. It catches that error alone: a parse failure, which is a plain ValueError, still names its case and writes nothing.
• 16 new cases:
  • RSS: co-core's trimmed Tailscale feed, verbatim, under text/xml and under application/rss+xml;
  • an Atom feed, a feed with no items (empty), a generic XML document, and an xpath on XML;
  • 10 refusals, covering both layers that refuse: extraction_config_from_spec and each extractor.
• Docs: the spec (with an amendment entry), AGENTS.md, and DEPLOYMENT.md's bump procedure.

The catch-up reading (0.19.7 → 0.32.0)

I checked these with git diff between the tags, not only the CHANGELOG:

• Unchanged: the bus (streams.py changed in comments only), the stores, canonical_text, spec_fingerprint, and the content.process / content.derived models (a changes.py docstring aside).
• No security fixes are recorded anywhere in the range.
• The contract change is resolve_dispatch_essence: .xml, .rss and .atom now map to the XML essences. Processor never re-resolves the essence, so whichever side moves first, a URL like that behind an ambiguous header stays on the HTML path, as it is today. Coordination is on watcher#366.

Bump note

Regenerated co-core-direct on 0.32.0: 16 moved, all added. None of the 13 existing goldens moved.

feed_atom, feed_no_items, feed_rss, feed_rss_rss_essence, xml_document, xml_xpath: added
refused_css_on_csv, refused_css_on_pdf, refused_css_on_xml, refused_empty_xpath,
refused_invalid_xpath, refused_jsonpath, refused_regex, refused_unknown_algorithm,
refused_xpath_on_csv, refused_xpath_on_html: added

• feed_rss = sha256:78e284978faf41fbe8ccd8dc06e6d2803d9788097fca633486dfbbbdc024be09. That is co-core's hand-pinned digest for the same fixture, and a test now holds it.
• The real corpus passes unchanged, 9/9 (all text/html + css).
• The whole corpus runs through the child under required containment (Landlock ABI 6): the XML extractor stays inside the allowlist, and every refusal comes back raised.

Verification

• uv run pytest: 650 passed, 14 skipped. Every skip is a secret that isn't readable here.
• uv run pytest -m integration: 56 passed.
• ruff check, ruff format --check and uv lock --locked are clean.

Review

Two rounds of review, with each fix in its own CR N commit:

• Round 1: CR 1, 2 and 6 correct the spec. CR 3 adds a pin test that holds build_wheelhouse.sh's default tag to the pin. CR 4 adds a test that the generator lists a case moving between a digest and a refusal (cannobserv#557 will move one). CR 5 names each layer's refusals in the coverage test, instead of counting them.
• Round 2: CR 9 names the default tag in the pin test's docstring.

Watcher coordination

Posted on watcher#366: the notice, with the new processor_version, the bump note and why either order is safe, and a correction: xpath on PDF / CSV is refused too.

After merge

1. Deploy once Watcher acknowledges on watcher#366, or once the operator rules that the acknowledgement isn't needed.
2. Run scripts/deploy.sh.
3. Live-spec gate. The journal doesn't record specs, but the same 4 info sources have run since 2026-10-05, with 204 of 204 commands complete. The first command from each must complete, and no extraction_error may carry UnsupportedExtractionError. If either fails, roll back.

🤖 Generated with Claude Code