fix(extract): respect content-visibility hidden subtrees

[ch-z-hc]

Bug

bsk extract includes text and links from subtrees hidden with CSS content-visibility: hidden. Discovery also returns tables and lists inside those subtrees, and explicit extraction accepts them instead of returning extract_target_hidden.

For example, extracting this table currently returns Visible\nHidden text:

<table id="t"><tr><td>Visible
  <div style="content-visibility: hidden">Hidden text</div>
</td></tr></table>

Chrome renders only Visible. The collector already excludes display-none and zero-opacity subtrees, but its cached ancestor check does not handle content visibility.

Fix

Add the missing rule to that existing ancestor check. Preserve rendered content on display types where Chromium ignores the property, and keep offscreen content-visibility: auto content extractable. No protocol, dependencies or page mutations are added.

Verification

• Final regression tests against unchanged main: 38 passed, 4 failed. With the fix: 42 passed, including all 11 real-Chrome extraction tests.
• Browser cases cover hidden text, links and containers, descendants trying to restore visibility, offscreen auto content, hidden cells, and ignored inline/contents/table display types.
• Default extension suite: 2,661 passed, 245 existing opt-in skips. Rechecked at head d87ca6798b647107a3c7bbf79902397e72e90269 with two workers; all 11 real-Chrome extraction tests also passed in a separate opt-in run.
• Extension TypeScript and production build passed again at the current head. Changed-file Biome, repository Biome with formatting disabled, Stylelint, DSH typecheck, localization tests (64) and VOM tests (82) passed.

The browser suite serializes the production collector into an isolated renderer world; it is not an installed-extension CLI end-to-end run. Other opt-in browser suites and Rust tests were not run. The DSH runtime suite was subsequently checked: 536 passed, 2 failed. An isolated checkout of unchanged base d9a78aed47874459ae1e44523a597e42dd9acf99, using the same dependencies and two workers, has the same 536 passed, 2 failed result. Both failures are the existing runner kill-grace cases (settles after the kill grace when neither exit nor close ever fires and settles killAll and killFor children that never report back); this PR does not change the DSH package. This does not claim the full pnpm lint command passed.

Implemented and tested with assistance from OpenAI Codex. No independent human review is claimed.