fix(wiki): re-ingest modified text files via content-hash change detection

mateclaw-server/src/main/java/vip/mate/wiki/service/WikiDirectoryScanService.java

@@ -137,32 +137,36 @@ public class WikiDirectoryScanService {
                 String fileName = file.getFileName().toString();
                 String ext = getExtension(fileName);
 
-                // 基于 sourcePath 去重
+                if (TEXT_EXTENSIONS.contains(ext)) {
+                    // Text files: dedup by content hash, so an unchanged file is
+                    // skipped while a modified file (new hash) is re-ingested.
+                    String content = Files.readString(file, StandardCharsets.UTF_8);
+                    boolean fresh = rawService.ingestTextFileFromScan(kbId, fileName, absolutePath, content);
+                    if (fresh) {
+                        added++;
+                    } else {
+                        skipped++;
+                    }
+                    continue;
+                }
+
+                // Binary files: dedup by source path. Re-ingest on content change
+                // is not detected here (hashing large binaries each scan is
+                // expensive) — modified binaries should be re-uploaded explicitly.
                 WikiRawMaterialEntity existing = rawService.findBySourcePath(kbId, absolutePath);
                 if (existing != null) {
                     skipped++;
                     continue;
                 }
-
+                String sourceType = switch (ext) {
-                if (TEXT_EXTENSIONS.contains(ext)) {
+                    case "pdf" -> "pdf";
-                    // 文本文件：读取内容；记录 source path 以便重复扫描去重
+                    case "docx", "doc" -> "docx";
-                    String content = Files.readString(file, StandardCharsets.UTF_8);
+                    case "pptx", "ppt" -> "pptx";
-                    WikiRawMaterialEntity textRaw = rawService.addText(kbId, fileName, content);
+                    case "xlsx", "xls" -> "xlsx";
-                    if (textRaw != null) {
+                    case "html", "htm" -> "html";
-                        rawService.updateSourcePath(textRaw.getId(), absolutePath);
+                    default -> "text";
-                    }
+                };
-                } else {
+                rawService.addFile(kbId, fileName, sourceType, absolutePath, Files.size(file));
-                    // 二进制文件：直接引用原始路径，不复制
-                    String sourceType = switch (ext) {
-                        case "pdf" -> "pdf";
-                        case "docx", "doc" -> "docx";
-                        case "pptx", "ppt" -> "pptx";
-                        case "xlsx", "xls" -> "xlsx";
-                        case "html", "htm" -> "html";
-                        default -> "text";
-                    };
-                    rawService.addFile(kbId, fileName, sourceType, absolutePath, Files.size(file));
-                }
                 added++;
 
             } catch (Exception e) {

mateclaw-server/src/main/java/vip/mate/wiki/service/WikiRawMaterialService.java

@@ -86,6 +86,32 @@ public class WikiRawMaterialService {
                         .eq(WikiRawMaterialEntity::getSourcePath, sourcePath));
     }
 
+    /**
+     * Import a text file discovered by a directory scan, detecting content
+     * changes by hash: unchanged content (a raw with the same hash already
+     * exists) is a no-op, while changed content creates a new raw and triggers
+     * processing — so a modified file is re-ingested rather than silently
+     * skipped. The originating path is recorded for diagnostics.
+     *
+     * @return {@code true} when the file was newly ingested (new or changed
+     *         content), {@code false} when skipped as unchanged
+     */
+    public boolean ingestTextFileFromScan(Long kbId, String fileName, String absolutePath, String content) {
+        String hash = computeHash(content);
+        WikiRawMaterialEntity sameContent = rawMapper.selectOne(
+                new LambdaQueryWrapper<WikiRawMaterialEntity>()
+                        .eq(WikiRawMaterialEntity::getKbId, kbId)
+                        .eq(WikiRawMaterialEntity::getContentHash, hash)
+                        .last("LIMIT 1"));
+        // addText dedups internally by hash, so this reuses sameContent when
+        // unchanged and inserts + triggers processing when the content differs.
+        WikiRawMaterialEntity raw = addText(kbId, fileName, content);
+        if (raw != null) {
+            updateSourcePath(raw.getId(), absolutePath);
+        }
+        return sameContent == null;
+    }
+
     /**
      * Record the originating file path on a raw material via a partial update,
      * so a later directory re-scan can dedup it by source path. Used for

mateclaw-server/src/test/java/vip/mate/wiki/service/WikiSourceWatcherServiceE2ETest.java

@@ -54,6 +54,14 @@ class WikiSourceWatcherServiceE2ETest {
         Files.writeString(sourceDir.resolve("note-c.md"), "# Note C\n\ncontent c");
         int secondAdded = watcherService.runScanCycle();
         assertEquals(1, secondAdded, "only the newly added file should ingest");
+
+        // Re-scanning with no changes ingests nothing.
+        assertEquals(0, watcherService.runScanCycle(), "unchanged files must not re-ingest");
+
+        // Modifying an existing file's content re-ingests it (content hash changed).
+        Files.writeString(sourceDir.resolve("note-a.md"), "# Note A\n\nEDITED content a");
+        int afterEdit = watcherService.runScanCycle();
+        assertEquals(1, afterEdit, "a modified file must be re-ingested");
     }
 
     @Test