feat(ai-core): trace instrumentation, LLM thinker, decay scheduler

- orchestrator: Tracer interface with spans for intent analysis, persona,
  subsession dispatch, context build, LLM streaming and each tool call;
  traces finish when the stream completes — the admin Trace waterfall
  finally receives data
- background: Thinker generates real LLM thoughts (post-chat + silence
  triggers, recent history as context, empty/overlong outputs discarded);
  thoughts broadcast to web clients as 'thinking' WS messages; pipeline
  postChatHook triggers post-chat thinking after each round
- memory: ApplyDecayAll + ListUserIDs with a 6h decay/purge cron in main
  (was implemented but never scheduled)

backend/services/memory/service.go

@@ -161,6 +161,23 @@ func (s *Service) ApplyDecay(ctx context.Context, userID string) (int, error) {
 	return updated, nil
 }
 
+// ApplyDecayAll applies decay for every user that has memories.
+func (s *Service) ApplyDecayAll(ctx context.Context) (int, error) {
+	users, err := s.repo.ListUserIDs(ctx)
+	if err != nil {
+		return 0, err
+	}
+	total := 0
+	for _, uid := range users {
+		n, err := s.ApplyDecay(ctx, uid)
+		if err != nil {
+			return total, err
+		}
+		total += n
+	}
+	return total, nil
+}
+
 // --- Keyword extraction (sliding window n-grams) ---
 
 // extractNGrams extracts all n-gram substrings from text.