feat(bizdev): 成功解析的材料一律写入 material_chunks

大小阈值只决定全文注入还是目录加检索片段，不再决定是否切块。上传后的异步解析同样跳过 60KB 门槛。

Co-authored-by: Cursor <cursoragent@cursor.com>

internal/bizdev/materials.go

@@ -17,7 +17,8 @@ import (
 
 // ResolveMaterials 将请求材料 + 项目 CAS 资产解析为可注入工作区的 squad 材料。
 // 项目内全部有效资产自动纳入（含私密文档）；显式附件优先并去重。
-// 已配置 MinerU 时优先使用解析缓存；大文档改为目录 + RAG 片段注入。
+// 已配置 MinerU 时优先使用解析缓存。解析成功的材料一律切块写入 material_chunks；
+// 正文不超过全文阈值时仍全文注入，超过阈值时改为目录 + RAG 片段注入。
 func (s *Service) ResolveMaterials(ctx context.Context, req TaskRunRequest) ([]squad.MaterialItem, map[string]any, error) {
 	manifest := map[string]any{
 		"explicit_materials": len(req.Materials),
@@ -173,6 +174,15 @@ func (s *Service) materialToItems(
 		base = "material.bin"
 	}
 
+	// 切块入库与注入方式分开：任何成功解析的正文都写入 material_chunks，失败只记日志。
+	if s.materialIndex != nil {
+		if err := s.materialIndex.IndexChunks(ctx, projectID, blobSHA, base, text); err != nil {
+			slog.Warn("material chunk index failed", "file", base, "project_id", projectID, "err", err)
+		} else {
+			meta["chunks_indexed"] = true
+		}
+	}
+
 	// 小文档：全文注入，文件名改为 .md 避免 workspace 二次按扩展名刮串冲掉页锚点
 	if len(text) <= fullMax {
 		outName := textMaterialName(base)
@@ -191,7 +201,6 @@ func (s *Service) materialToItems(
 	toc := materialindex.BuildTOC(base, text)
 	ragBody := ""
 	if s.materialIndex != nil {
-		_ = s.materialIndex.IndexChunks(ctx, projectID, blobSHA, base, text)
 		hits, searchErr := s.materialIndex.Search(ctx, projectID, blobSHA, ragQuery, 0)
 		if searchErr == nil {
 			ragBody = materialindex.FormatRAGMarkdown(base, hits)

internal/bizdev/materials_index_test.go

@@ -0,0 +1,45 @@
+package bizdev
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"chat-agent-system/internal/config"
+	"chat-agent-system/internal/materialindex"
+)
+
+func TestMaterialToItems_SmallDocStillFulltextAfterIndex(t *testing.T) {
+	svc := &Service{materialIndex: materialindex.NewPipeline(&config.Config{}, nil)}
+	items, meta, err := svc.materialToItems(context.Background(), "note.txt", "abc123", []byte("能量参数是 2.5 bar"), 0, 0, 7, "能量")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if meta["inject_mode"] != "fulltext" {
+		t.Fatalf("small doc inject_mode=%v", meta["inject_mode"])
+	}
+	if meta["chunks_indexed"] != true {
+		t.Fatalf("small doc must attempt chunk index, meta=%v", meta)
+	}
+	if len(items) != 1 || !strings.Contains(string(items[0].Content), "能量参数是 2.5 bar") {
+		t.Fatalf("full text missing: %+v", items)
+	}
+}
+
+func TestMaterialToItems_LargeDocIndexesBeforeRAGInject(t *testing.T) {
+	svc := &Service{materialIndex: materialindex.NewPipeline(&config.Config{MaterialFulltextMaxBytes: 8}, nil)}
+	body := []byte("能量参数是 2.5 bar，并且还有更多正文用来越过全文阈值。")
+	items, meta, err := svc.materialToItems(context.Background(), "note.txt", "def456", body, 0, 0, 7, "能量")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if meta["chunks_indexed"] != true {
+		t.Fatalf("large doc must attempt chunk index, meta=%v", meta)
+	}
+	if meta["inject_mode"] != "rag" && meta["inject_mode"] != "rag_degraded" {
+		t.Fatalf("large doc must not use fulltext, inject_mode=%v", meta["inject_mode"])
+	}
+	if len(items) != 2 {
+		t.Fatalf("want toc+rag items, got %d", len(items))
+	}
+}

internal/materialindex/pipeline.go

@@ -284,10 +284,8 @@ func (p *Pipeline) EnqueueAsync(fileName string, blobSHA string, data []byte, pr
 		if err != nil || !ok || strings.TrimSpace(md) == "" {
 			return
 		}
-		if utf8.RuneCountInString(md) > 0 && len(md) > p.FulltextMaxBytes() {
-			if idxErr := p.IndexChunks(ctx, projectID, blobSHA, fileName, md); idxErr != nil {
-				slog.Warn("materialindex index chunks failed", "sha", blobSHA[:min(12, len(blobSHA))], "error", idxErr)
-			}
+		if idxErr := p.IndexChunks(ctx, projectID, blobSHA, fileName, md); idxErr != nil {
+			slog.Warn("materialindex index chunks failed", "sha", blobSHA[:min(12, len(blobSHA))], "error", idxErr)
 		}
 	}()
 	return true