[Moskize91]

Page refinement could reject a valid current-page result because a neighbour's initial or adopted structure disagreed with it. The citation-on reproduction failed when a continuation named a publication cue that the previous page had excluded.

This refactors acceptance to validate only the complete current-page result. Source decoding discards initial semantic decisions; cross-page assembly mechanically applies current-page append declarations using output cursors, preserving unbound content when no cursor exists. It removes neighbour membership checks, successor trial adoption, cross-page metadata gates, and post-repair rebinding. Batch and streaming paths share assembly. PCEX v6 storage is unchanged.

The repair prompt separates observations from initial proposals and leaves JEV scores at admission. Regression coverage includes forbidden-evidence reads, independent page jobs, internal-invalid results, unbound streams, and transactional delivery. Smoke expectations now use stable source IDs and check the four actual printed footnotes and their continuation instead of requiring an artificial fifth definition.

Validation:

• Full suite: 1965 tests, OK, 6 skipped; independently repeated by the VGE reviewer.
• Pyright: 0 errors/warnings; pylint: 10.00/10.
• Offline replay of original citation-on replies and the six-page repair case passed.
• Final paid citation_footnotes gate: automated_passed=true, failures=[], intact printed definitions and cross-page continuation, no body/note leakage, preserved formula pixels. Three-page source/output previews inspected.

Earlier paid trials exposed model classification errors and an obsolete definition-count assertion; those failed reports were retained. The final targeted run passed; this does not establish deterministic model accuracy.

Related to #616, #619, #620 and oomol/pdf-craft-service#314. Implements VGE issue 983.