feat: add bounded npc memory retention

docs/LEARNING_ROADMAP.md

@@ -684,10 +684,25 @@ The first communicated-knowledge slice is complete:
   malformed-provenance rejection, event-history invariance, save/checksum, and
   informed-choice regressions cover the complete slice.
 
+The first importance/retention slice is complete:
+
+- acquisition ticks order facts by when an NPC learned them rather than by the
+  objective event ID;
+- a current relationship cause is lasting, while each NPC keeps at most three
+  other recent facts;
+- a deterministic daily-sized review forgets recent facts at least one day old
+  without deleting objective event history or rolling back trust;
+- communication prefers lasting direct evidence, and a source-method snapshot
+  keeps listener provenance valid after the teller forgets;
+- the inspector shows lasting/recent counts, while schema-v7 migration,
+  boundary, capacity, cause-protection, save/checksum, and informed-choice
+  regressions cover the slice.
+
 This advances the first Milestone 6 evidence-to-choice path without claiming
-full memory or belief simulation. The practical next slice is importance and
-retention for known facts before expanding into distortion, secrecy, rumours,
-or opportunity generation.
+full belief simulation. The practical next sequence is a compact retained
+memory/person-history view and a visible social-reaction cue, then one bounded
+Milestone 7 rumour/opportunity proof without generalized distortion or
+free-form dialogue.
 
 Recently completed: