feat(knowledge): broaden context resolution and prompt query

- Walk knowledge group children recursively so selecting a parent group
  pulls snippets from its descendants too.
- Filter knowledge items to status='completed' across the prompt and
  active-snippet queries to avoid leaking in-progress documents.
- Dedupe vector candidates by point id and normalized text, then pick
  rerank+vector results with a per-item cap so prompts stay diverse.
- Tag each rendered snippet with a [Kn] reference id for prompt citing.
- Build the rule knowledge query from task params (title/keywords/etc.)
  before falling back to rule prompt content so retrieval matches the
  user's actual ask.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

server/internal/tenant/app/prompt_generate_service.go

@@ -657,16 +657,35 @@ func buildPromptRuleKnowledgeQuery(rule *promptRuleGenerationRecord, params map[
 		if rule.Scene != nil && strings.TrimSpace(*rule.Scene) != "" {
 			parts = append(parts, strings.TrimSpace(*rule.Scene))
 		}
-		if text := strings.TrimSpace(rule.PromptContent); text != "" {
-			parts = append(parts, text)
-		}
 	}
 	if taskName := strings.TrimSpace(extractString(params, "task_name")); taskName != "" {
 		parts = append(parts, taskName)
 	}
+	for _, key := range []string{
+		"title",
+		"topic",
+		"subject",
+		"brand_name",
+		"product_name",
+		"primary_keyword",
+		"key_points",
+		"extra_requirements",
+	} {
+		if text := strings.TrimSpace(extractString(params, key)); text != "" {
+			parts = append(parts, text)
+		}
+	}
+	if keywords := extractStringList(params["keywords"], 16); len(keywords) > 0 {
+		parts = append(parts, strings.Join(keywords, "\n"))
+	}
 	if target := strings.TrimSpace(extractString(params, "target_platform")); target != "" {
 		parts = append(parts, target)
 	}
+	if rule != nil {
+		if text := strings.TrimSpace(rule.PromptContent); text != "" {
+			parts = append(parts, text)
+		}
+	}
 	return strings.Join(parts, "\n")
 }