docs(ai): Revert-per-iteration shipped in Workbench

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

docs/architecture/COMPANION_BRAIN_ARCHITECTURE.md

@@ -1780,6 +1780,10 @@ Code:
     einmal pro Bucket statt pro Event
   - Filter: Mission (per query-string) + Modul (dropdown), Deep-Link ins
     Modul pro Event
+  - **Revert-per-Iteration**: Button pro Bucket, `data/ai/revert/`
+    Registry mit Inverse-Ops für TaskCreated/Completed, CalendarEvent-
+    Created, PlaceCreated, DrinkLogged. Newest-first Reihenfolge,
+    RevertStats-Summary ("X zurückgenommen · Y nicht unterstützt").
 - [x] Schritt 7a — System-Actor-Wrapping für Projections (streaks-Tracker)
 - [x] Schritt 8 — mana-sync Go + Postgres-Migration für `actor`-Feld
   - `sync_changes.actor JSONB` Column (idempotent `ADD COLUMN IF NOT EXISTS`)

services/mana-ai/CLAUDE.md

@@ -21,7 +21,6 @@ What works end-to-end:
 
 Intentionally **not yet** implemented (future work):
 - [ ] Materialized mission snapshot — full LWW replay per tick is O(N changes). Fine pre-launch; revisit when user count grows.
-- [ ] Revert-per-iteration in the Workbench timeline — would undo the AI's approved writes. Needs per-event-type inverse logic (delete reverts create, uncomplete reverts TaskCompleted, …).
 
 ## Port: 3066