feat(knowledge): multi-query retrieval with LLM query rewrite

Add a shared structured query builder (knowledge_query.go) and switch
all generation paths to it, so retrieval queries carry task intent
(type, brand, region, keywords, key points) instead of a raw prompt.

ResolveContext now rewrites the query into multiple sub-questions via
the URL-markdown model, embeds and searches each, and merges/dedupes
candidates before rerank; falls back to parsed fallback questions when
rewrite is unavailable or returns too few. Resolve logs gain query
list, count, and rewrite stage/model for observability.

- knowledge_query.go/_test.go: BuildKnowledgeQuery + intent normalization
- knowledge_service.go: per-query embed/search loop, query rewrite, logs
- template_prompt/template_service/prompt_generate/article_imitation/
  kol_generation_worker: adopt the shared query builder
- generation_observability: richer task error logging

server/internal/tenant/app/template_service.go

@@ -491,7 +491,7 @@ func (s *TemplateService) executeGeneration(ctx context.Context, job generationJ
 			ctx,
 			job.TenantID,
 			knowledgeGroupIDs,
-			buildGenerationKnowledgeQuery(job.Params),
+			buildGenerationKnowledgeQuery(job.TemplateKey, job.TemplateName, job.Params),
 		)
 		if resolveErr != nil {
 			s.failGeneration(ctx, job, title, "knowledge_resolve", resolveErr)
@@ -604,7 +604,7 @@ func (s *TemplateService) failGeneration(ctx context.Context, job generationJob,
 	now := time.Now()
 	logCtx := generationJobLogContext(job, stage, "", "failed", GenerationTaskResultFailure)
 	RecordGenerationTaskEvent(GenerationTaskEventFailure, logCtx)
-	LogGenerationTaskWarn(s.logger, "template generation failed", logCtx, zap.Error(genErr))
+	LogGenerationTaskWarn(s.logger, "template generation failed", logCtx, GenerationTaskErrorFields(genErr)...)
 
 	auditRepo := repository.NewAuditRepository(s.pool)
 	articleRepo := repository.NewArticleRepository(s.pool)