Scope whole-memo claim-quote exemptions to the claims table

[samrusani]

Summary

• The in-process language backstop now applies claim-quote exemptions only inside the rendered claims-table region (## Claims vs evidence through the next ## heading).
• Render-only reruns can no longer treat a hand-edited short claim, or a stored claim phrase copied into narrative, as a memo-wide whitelist.
• Identity exemptions, quotation verification, the per-field gate, and the CI memo scan are unchanged. Missing start or closer fails closed.

Test plan

• Four original exemption-attack shapes still fail closed
• Hand-tamper: claim fraud plus narrative fraud on rerun --render-only exits 1 and writes no memo
• Hand-tamper: full puffery phrase copied into narrative fails and leaves the old memo untouched
• Short claim fraud with clean prose still renders in the table
• Golden memo still passes the backstop
• Five gates: 774 tests, ruff, format, mypy, language check

[greptile-apps]

Greptile Summary

The PR narrows stored-claim language exemptions to the rendered claims-table region while retaining memo-wide identity exemptions and fail-closed behavior. The new delimiter parser can, however, mistake unrestricted rendered content for a structural heading.

• Adds a region-aware whole-memo language scanner.
• Routes the pipeline backstop through the new scanner.
• Adds render-only tampering and golden-memo regression coverage.
• Updates architecture, decision, and handoff documentation.

Confidence Score: 4/5

The boundary parsing defect should be fixed before merging because valid claim quotations can be rejected when rendered fields contain heading-like lines.

The scanner derives structural boundaries from the final rendered text even though render-only casefiles permit multiline strings, allowing field content to move the closer or start and exclude part of the claims table from its required exemption.

Files Needing Attention: src/coldscreen/language.py

Important Files Changed

Filename	Overview
src/coldscreen/language.py	Introduces region-scoped memo scanning, but content-derived heading boundaries can truncate or shift the claims region.
src/coldscreen/pipeline.py	Correctly routes the whole-memo backstop through the new region-aware helper.
tests/test_language.py	Covers missing, unclosed, repeated, and ordinary table boundaries but not headings injected through rendered fields.
tests/test_exemption_attacks.py	Adds useful render-only tampering regressions for memo-wide claim exemption attacks.
tests/test_golden.py	Confirms both the committed and freshly rendered golden memos pass the scoped backstop.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Load casefile] --> B[Render memo]
  B --> C[Find first Claims vs evidence heading]
  C --> D[Find next line beginning with ##]
  D --> E[Apply claim and identity exemptions inside region]
  E --> F[Apply identity exemptions outside region]
  F --> G{Banned terms found?}
  G -->|Yes| H[Reject memo write]
  G -->|No| I[Write memo]

Loading

Prompt To Fix All With AI

### Issue 1
src/coldscreen/language.py:212-216
**Content corrupts table boundaries**

When a render-only casefile includes an embedded `## ` line in a claim-table field, or an exact `## Claims vs evidence` line in earlier prose, `_claims_table_region` treats that content as a structural delimiter. This excludes legitimate claim rows from the exemption region and causes the language backstop to reject the rendered memo.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (1): Last reviewed commit: "Scope whole-memo claim-quote exemptions ..." | Re-trigger Greptile}

[greptile-apps]

Content corrupts table boundaries

When a render-only casefile includes an embedded ## line in a claim-table field, or an exact ## Claims vs evidence line in earlier prose, _claims_table_region treats that content as a structural delimiter. This excludes legitimate claim rows from the exemption region and causes the language backstop to reject the rendered memo.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/coldscreen/language.py
Line: 212-216

Comment:
**Content corrupts table boundaries**

When a render-only casefile includes an embedded `## ` line in a claim-table field, or an exact `## Claims vs evidence` line in earlier prose, `_claims_table_region` treats that content as a structural delimiter. This excludes legitimate claim rows from the exemption region and causes the language backstop to reject the rendered memo.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.