Preserve material distinctions in rendered stories

[R055LE]

Summary

Closes #115. Stacked on #113 (issue-110-model-evaluation), which depends on #109.

• Review the actual public render for material omissions with bounded, source-grounded diagnostics.
• Judge implied relationships across the narrative, including distinct experimental settings.
• Keep each item's exact citation support independent from broader completeness evidence.
• Reject unchanged failed items and hidden-only repairs; preserve one revision and four calls maximum.
• Share review payload/parsing between production and evaluation. Explicitly preserve required citations through revision.

Validation

• Full suite: 634 passed, 87.38% coverage; affected suite 112 passed.
• Final compact revision wording passed focused context/revision tests, Ruff and diff checks.
• Independent code review: no blockers.
• Final-head CI passed on Python 3.12 and 3.14: https://github.com/R055LE/roger/actions/runs/38474683818
• Live reviewer calibration: 6/6 exact authored-gate matches. All six complete request records remain byte-identical at this final source.

Integrated acceptance remains incomplete

The final held-out render preserves the material experimental distinctions and caveats. Both reviews pass the completeness gate. However, the complete workflow is rejected: revision replaces the accepted stage citation with one describing only construction of the testbed while retaining stage: demonstrated. Final support review and independent assessment reject that item. The remaining grounded revision/stage compliance belongs to #107.

Earlier false acceptance, the402 interruption, and the empty-citation revision remain preserved. No accepted workflow or human editorial acceptance is claimed. The optional marketing workflow repeat was not run; its material-omission controls passed calibration.

#115 used 20 attempts/$0.5201765 known plus $0.74 uncertain reserve. Full study: 196/196 attempts, $3.06141310 known plus $1.04 retained uncertainty. Further paid testing needs a new call allowance. Provider-credit restoration succeeded; funding is no longer the blocker.

Limits

Worst-case rejected revision is 17,994/18,000 characters. An all-eight maximum-length-source combine can still exceed the unchanged review ceiling and fail closed without truncation. Repository regressions are synthetic reproductions; captured originals remain immutable.

Draft pending integrated acceptance and dependent PRs. Human editorial and Scout acquisition acceptance remain separate. No production activation or publication.