Exercise IB document extraction on the hostile corpus

[isomorphisms]

Advances #43.

Pins isomorphisms/ai-ci@c1e423d506f6adfc8c1c9659a4faaf94ae84eaf0 and runs the Idriç candidate path against the canonical valid and malformed HTML fixtures.

This revision makes the chosen recovery boundary explicit: IB.RecoveredInformation sits between tokenization and InformationView and implements the first declared useful-document contract, document_log_subset_v0. The recovered log carries source-fixture identity and ordered tag/text events, with explicit implied-end events rather than pretending to provide a browser DOM or full WHATWG recovery.

The malformed fixture now has to exercise that contract before document construction can pass: it requires implied closure for malformed paragraphs, table cell/row structure, and the unterminated link at EOF. The resulting InformationView must retain the recovered A/B row and downstream extraction must recover syllabus at /syllabus. Both valid and malformed local-fixture receipts are therefore expected to finish with first_failure=none.

The receipt still truthfully marks network as SKIP/not_applicable_local_fixture. IB reads the exact pinned corpus bytes locally; it does not invoke the external oracle or use it as fallback. Network ownership remains in the coordinated ICU work, not IB.