[andyhsu6]

Closes #288.

Problem

Character profiles, world-setting facts and sync markers advance by chapter number but never roll back when chapters are deleted. On a real project, after removing chapters 58–70:

• character state_updated_chapter / profile_synced_through_chapter still said 62;
• profile_facts still cited chapters 58/59/61/62;
• 612 world-setting claims, 118 of them sourced only from deleted chapters, still fed the world-building text that goes into every chapter prompt.

Change

• Delete-time clamp (app/services/derived_state.py, called after DELETE /chapters/{id}):
  • drop character facts and world-setting claims whose source chapters are gone;
  • clear out-of-range current_state and reset its marker to NULL (setting it to the cutoff would make the update service skip re-analysis of earlier chapters — the anti-regression check is chapter < state_updated_chapter);
  • clamp profile_synced_through_chapter and world_profile_synced_through_chapter.
• Cutoff = last chapter with content, not the last chapter row. Batch generation creates empty chapter shells; using shell numbers would keep facts from deleted chapters legal. This was caught while verifying: with the row-based cutoff, 118 stale claims survived.
• Read-time guard: the 1-N context renderer skips state whose source chapter is after the chapter being written.
• Consistency check: GET /project/{id}/consistency reports stale facts/markers (read-only, truncated list + total), so this drift is detectable rather than silent.

Verification

real project: cutoff=58, dropped_world_facts=118, remaining stale=0, ok=true
.venv/bin/python -m pytest tests/test_derived_state_cutoff.py -q  → 6 passed
.venv/bin/python -m pytest tests/ -q -p no:cacheprovider          → 1665 passed

Tests cover: clamping facts/state/markers, world-claim pruning (evidence kept when partly in range, claims without evidence preserved), idempotency, "clean data produces no UPDATE", and the read-time skip for both directions.

Independent review

Reviewed by a different model. One MAJOR was raised — world-setting facts were only watermark-clamped while their content stayed — and is addressed here (pruning + re-render avoided on purpose: re-rendering the four text columns from facts would rewrite existing content, verified by length comparison). Reviewer MINORs also handled: marker reset to NULL, unparseable facts counted in the report, endpoint list truncated to 20 with a total, and the no-op write path now has a test.

Not done here (deliberately): soft delete and an archive/trash table. If chapters should be recoverable, that is a separate change with its own read-path audit.

Desensitized per repository policy: no source text, titles or character names are included.