Exclude appended selected excerpts from trusted .txt snippet-cache coverage #1362

Finding

scripts/evidence_snippet_audit.py:207-210 treats every .txt cache as independently retrieved source text and returns before the selected-excerpt exclusions used for Markdown caches. Real PubMed abstracts with appended curator-selected excerpts therefore certify those excerpts as source coverage.

Reproduced read-only using the production cache_text("PMID:42613016") function at the published batch25 tree (commit 5f7e0faecd6486addf38d462234f7cb6067d29a4; relevant code and cache unchanged by the website integration):

trusted: True
SELECTED PUBLIC PMC ARTICLE EXCERPTS marker in trusted text: True
selected-only phrase in trusted text: True

The selected-only phrase was The initial inoculation ratio was set to 1:1 in model simulations. It occurs after the explicit selected-excerpts marker in references_cache/PMID_42613016.txt, not in its independently retrieved PubMed abstract. The issue is provenance, not a finding that the biological statement is false. No upstream reference-validator end-to-end reproduction or exhaustive affected-snippet count is claimed here.

A narrow ignored-inclusive scan of references_cache/*.txt also finds PMID_41381587.txt with SELECTED FULL-TEXT EXCERPTS FROM PMC12816627. These two marker matches are not an exhaustive inventory of all possible curated-excerpt formats.

Acceptance

• Preserve the independently retrieved PubMed abstract, but exclude appended selected-excerpt sections from the snippet auditor's trusted source text.
• Cover both observed markers and genuine unmodified abstract/full-text caches with focused regression tests; retain the Markdown selected-excerpt guard from Do not promote selected-excerpt caches to independent snippet verification #1237.
• Re-audit all records referencing affected caches and report newly exposed source-coverage gaps honestly. Do not repair them by copying curated snippets back into trusted cache content or by weakening matching.
• Keep historical review outputs as provenance, and distinguish primary-source reading from independent local cache verification in new review outputs.

Relationship To Earlier Work

Closed #1237 addresses Markdown content_type: selected_excerpts; this is the earlier .txt return path, which that fix does not reach. Closed #592 concerns a related circular-validation defect in snippet-listing caches. Closed #1019-#1022 and #1028 added selected excerpts for DCPP/RAMC records; they did not establish independent cache provenance. Closed #1147 concerns a full-text retrieval sentinel, not this audit trust decision.

Deduplication checked all 696 current open/closed issue titles and bodies plus ignored-inclusive scoped local reports, history, research, tests, and scripts. Every issue comment was not searched. Identified during preparation for the next causal-graph batch; not part of PR #1361, and not fixed by its graph-only changes.