Memo references and appendix boundaries

This independent source review follows two defects found by the complete-corpus outlier check. It adds regression controls without changing earlier experiments, source captures or published projections. The source expectations and generic repair were developed concurrently; this was not a blinded or preregistered experiment.

Source findings

• ADM memo 06/21, PDF page 5: C2130 begins an extracted line, but continues see ADM inside memo paragraph 11. It is a reference, not the memo's own numbering. The actual annotation list is separate, on page 21. The rendered source page shows the paragraph and reference.
• DMG volume 8 amendment 27, PDF pages 107–109: paragraph 44650 includes Example 1 on page 107 and the complete Example 2 on page 108. A separate Appendix 3 starts on page 109. The historical PDF's extracted heading is A p p e n di x 3. Its spacing is an extraction artefact, not a reason to attach the appendix to paragraph 44650.

The source fixtures bind the original PDF, extraction and inventory hashes, exact source spans and expected limits. They retain the historical amendment role. The images were rendered from those frozen PDFs with pdftoppm, at a maximum dimension of 1,100 pixels, and visually inspected. They are source-review aids under the repository's source notice.

Retained checks

The previous-parser observation rejects all three final controls. The repaired-parser observation passes all three:

1. The wrapped memo citation never becomes a main paragraph.
2. Actual source-declared memo sections and final annotation references remain separate, with legal dependency status unresolved.
3. The appendix starts a new section after the full continuation example, while every original source byte remains accounted for.

The implementation review binds the reviewed code, source-tree inputs and tests. The repair is generic: memo inventory roles govern the numbering distinction; standalone appendix literals create boundaries. It does not contain these document identifiers or paragraph numbers.

An initial reviewer expectation failed: the test expected the visible “Past presence test” topic to be an admitted PDF heading tag. That was not supported by the retained tree alignment. The initial test is preserved. The corrected control uses the genuinely admitted Introduction and Background headings. This correction does not establish fine-grained memo structure: the Background section still spans pages 2–8, and its complete legal dependencies remain unknown.

Reproduce the bounded controls

uv run --locked python -m unittest discover -s scripts -p 'test_manual_memo_appendix_sources.py' -v

These tests read frozen local evidence and replay existing PDF structure. They do not acquire source documents, run Poppler, rebuild the corpus or call a model. Exact byte coverage, source-declared headings and passing boundary controls do not establish legal applicability, answerability or specialist acceptance.

The artefact manifest binds this follow-on evidence; the earlier auxiliary and wrapped-reference manifests remain unchanged.