fix(identity): 识别缓存不把一次指定扩散给不相容的同名资源,召回详情失败不写负缓存(H2、M8)

回填只认用户本人的指定,资源没有年份、或作品年份 / 类型与资源不相容时不回填;
读缓存时核对条目与资源的年份(±1,剧集允许更晚)与类型,不相容按未命中处理,
命中后的唯一候选仍经 resolveIdentity 按资源年份 / 类型打分。
TMDB 召回中有候选拉详情失败时,这一轮无匹配不算结论,不写负缓存。

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Cxwgnn2pbzQ7rXv3rq7fei

apps/server/src/modules/sources-identity/service/identity-admin.ts

@@ -226,7 +226,7 @@ export async function assignReleaseIdentity(
         });
         return { mediaEntityId: entityId, contentUnitIds: unitIds };
       });
-      await backfillRecognizeCache(rt, parsed, done.mediaEntityId);
+      await backfillRecognizeCache(rt, parsed, done.mediaEntityId, actor.type);
       if (created) await enqueueNewEntityRefresh(ctx, done.mediaEntityId);
       return done;
     } catch (e) {

apps/server/src/modules/sources-identity/service/resolve.ts

@@ -196,24 +196,31 @@ async function tmdbCandidates(
   parsed: ParsedRelease,
   useCache: boolean,
   siteIds: ReturnType<typeof siteExplicitIds> = [],
-): Promise<{ remote: MetadataCandidate[]; cacheHit: boolean; meta: RecognizeMeta | null }> {
+): Promise<{ remote: MetadataCandidate[]; cacheHit: boolean; meta: RecognizeMeta | null; detailFailed: boolean }> {
   const tmdb = await rt.tmdb();
-  if (!tmdb) return { remote: [], cacheHit: false, meta: null };
+  if (!tmdb) return { remote: [], cacheHit: false, meta: null, detailFailed: false };
   const meta = recognizeMetaOf(parsed);
   // 站点详情页给出的 TMDB / IMDb ID:按 ID 直取(TMDB 详情 / find),命中就不再按标题搜索
   const byId = await tmdbByExplicitIds(tmdb, siteIds);
-  if (byId.length) return { remote: byId, cacheHit: false, meta };
+  if (byId.length) return { remote: byId, cacheHit: false, meta, detailFailed: false };
   if (meta && useCache) {
-    const hit = await loadRecognizeCache(rt, meta);
+    // 命中条件含年份 / 类型相容(loadRecognizeCache);命中后的唯一候选照样经 resolveIdentity 按资源年份 / 类型打分
+    const hit = await loadRecognizeCache(rt, meta, parsed);
     if (hit) {
       if (hit.id && hit.title && (hit.type === "movie" || hit.type === "tv")) {
-        return { remote: [await tmdb.details(hit.type, hit.id)], cacheHit: true, meta };
+        return { remote: [await tmdb.details(hit.type, hit.id)], cacheHit: true, meta, detailFailed: false };
       }
-      return { remote: [], cacheHit: true, meta };
+      return { remote: [], cacheHit: true, meta, detailFailed: false };
     }
   }
-  const remote = await recallTmdbCandidates(tmdb, tmdbRecallInput(parsed));
+  let detailFailed = false;
-  return { remote, cacheHit: false, meta };
+  const remote = await recallTmdbCandidates(tmdb, {
+    ...tmdbRecallInput(parsed),
+    onDetailError: () => {
+      detailFailed = true;
+    },
+  });
+  return { remote, cacheHit: false, meta, detailFailed };
 }
 
 async function tmdbByExplicitIds(
@@ -494,7 +501,11 @@ export async function resolveRelease(rt: SourcesRuntime, input: ResolveInput): P
   }
   let remote: MetadataCandidate[] = [];
   let lookupError: string | null = null;
-  let recognize: { cacheHit: boolean; meta: RecognizeMeta | null } = { cacheHit: false, meta: null };
+  let recognize: { cacheHit: boolean; meta: RecognizeMeta | null; detailFailed: boolean } = {
+    cacheHit: false,
+    meta: null,
+    detailFailed: false,
+  };
   try {
     const r = await tmdbCandidates(rt, parsed, !input.force, siteIds);
     remote = r.remote;
@@ -512,7 +523,8 @@ export async function resolveRelease(rt: SourcesRuntime, input: ResolveInput): P
       top?.provider === "tmdb"
         ? remote.find((m) => m.externalId === top.externalId && m.entityType === top.entityType)
         : undefined;
-    const info = hit ?? (resolution.status === "no_match" ? null : undefined);
+    // 召回时有候选拉详情失败(缺别名 / 季年份):这一轮"无匹配"不算结论,不写负缓存
+    const info = hit ?? (resolution.status === "no_match" && !recognize.detailFailed ? null : undefined);
     await saveRecognizeCache(rt, recognize.meta, info);
   }
 
@@ -1136,7 +1148,11 @@ export async function identifyParsed(
   const lookup = await lookupReason(rt, parsed, trigger, local.length > 0, null);
   let remote: MetadataCandidate[] = [];
   let lookupError: string | null = null;
-  let recognize: { cacheHit: boolean; meta: RecognizeMeta | null } = { cacheHit: false, meta: null };
+  let recognize: { cacheHit: boolean; meta: RecognizeMeta | null; detailFailed: boolean } = {
+    cacheHit: false,
+    meta: null,
+    detailFailed: false,
+  };
   if (lookup) {
     try {
       const r = await tmdbCandidates(rt, parsed, true);
@@ -1154,7 +1170,8 @@ export async function identifyParsed(
       top?.provider === "tmdb"
         ? remote.find((m) => m.externalId === top.externalId && m.entityType === top.entityType)
         : undefined;
-    const info = hit ?? (resolution.status === "no_match" ? null : undefined);
+    // 召回时有候选拉详情失败(缺别名 / 季年份):这一轮"无匹配"不算结论,不写负缓存
+    const info = hit ?? (resolution.status === "no_match" && !recognize.detailFailed ? null : undefined);
     await saveRecognizeCache(rt, recognize.meta, info);
   }
   const candidates = resolution.ranked.slice(0, 5).map((r) => {

apps/server/src/modules/sources-identity/service/tmdb-recognize-cache.ts

@@ -65,8 +65,31 @@ function localeOf(tmdb: unknown): string {
   return (tmdb as { language?: string } | null)?.language ?? "zh-CN";
 }
 
-/** TmdbCache.get:未命中返回 null;类型冲突的脏条目删除后按未命中处理 */
+/** 资源是否带剧集线索(季 / 集 / 类型提示) */
-export async function loadRecognizeCache(rt: SourcesRuntime, meta: RecognizeMeta): Promise<RecognizeCacheEntry | null> {
+const tvLike = (p: ParsedRelease) => p.mediaTypeHint === "tv" || p.season !== undefined || p.episodes !== undefined;
+
+/**
+ * 缓存条目与资源自身的年份 / 类型是否相容(H2):资源有年份时条目年份须在 ±1 内(剧集允许资源年份更晚,后续季);
+ * 资源有剧集线索时条目须是剧集。负缓存不带年份类型,总是相容。
+ */
+export function entryCompatible(p: ParsedRelease, e: Pick<RecognizeCacheEntry, "id" | "type" | "year">): boolean {
+  if (!e.id) return true;
+  if (tvLike(p) && e.type === "movie") return false;
+  const y = e.year ? Number(e.year) : Number.NaN;
+  if (p.year === undefined || !Number.isFinite(y)) return true;
+  if (Math.abs(p.year - y) <= 1) return true;
+  return e.type === "tv" && p.year > y;
+}
+
+/**
+ * TmdbCache.get:未命中返回 null;类型冲突的脏条目删除后按未命中处理。
+ * 传入资源解析时再核对条目与资源的年份 / 类型,不相容按未命中处理(不删除,别的资源可能正好相容)。
+ */
+export async function loadRecognizeCache(
+  rt: SourcesRuntime,
+  meta: RecognizeMeta,
+  parsed?: ParsedRelease,
+): Promise<RecognizeCacheEntry | null> {
   const tmdb = await rt.tmdb();
   const key = recognizeKey(meta, localeOf(tmdb));
   const cache = sharedCache(rt.ctx);
@@ -77,6 +100,7 @@ export async function loadRecognizeCache(rt: SourcesRuntime, meta: RecognizeMeta
     await cache.delete(CACHE_REGIONS.tmdbRecognize, key);
     return null;
   }
+  if (parsed && !entryCompatible(parsed, v)) return null;
   return v;
 }
 
@@ -121,13 +145,17 @@ export async function saveRecognizeCache(
 /**
  * 人工认定后回填识别缓存(MP update_recognize_cache:覆盖名称负缓存,避免同名资源再次回查)。
  * 作品没有已接受的 TMDB 身份时不写。失败只记日志。
+ * H2:一次指定不扩散给不相容的同名资源——只回填用户本人的指定(Agent / 系统的不回填);资源没有年份不回填
+ * (键里年份是 None,会被所有无年份的同名资源命中);作品年份 / 类型与资源解析不相容(指定到了别的版本)不回填。
  */
 export async function backfillRecognizeCache(
   rt: SourcesRuntime,
   parsed: ParsedRelease,
   entityId: string,
+  actorType: string = "user",
 ): Promise<void> {
   try {
+    if (actorType !== "user" || parsed.year === undefined) return;
     const meta = recognizeMetaOf(parsed);
     if (!meta || !(await rt.tmdb())) return;
     const row = await rt.ctx.db
@@ -139,6 +167,12 @@ export async function backfillRecognizeCache(
       .where("i.status", "=", "accepted")
       .executeTakeFirst();
     if (!row || (row.entity_type !== "movie" && row.entity_type !== "tv")) return;
+    const entry = {
+      id: Number(row.external_id),
+      type: row.entity_type,
+      year: row.year === null ? null : String(row.year),
+    };
+    if (!entryCompatible(parsed, entry)) return;
     const md = (row.metadata ?? {}) as Record<string, { value?: unknown } | undefined>;
     const str = (v: unknown) => (typeof v === "string" && v ? v : null);
     await saveRecognizeCache(rt, meta, {

packages/adapters/src/tmdb/recall.ts

@@ -22,6 +22,8 @@ export interface TmdbRecallInput {
   maxDetails?: number;
   /** 候选池上限,默认 20 */
   limit?: number;
+  /** 拉详情失败(保留搜索结果继续)时回调:调用方据此知道这一轮召回不完整,例如不写负缓存 */
+  onDetailError?: (e: TmdbError, c: MetadataCandidate) => void;
 }
 
 const yearDistance = (c: MetadataCandidate, year: number | undefined, type: MediaEntityType): number => {
@@ -65,6 +67,7 @@ export async function recallTmdbCandidates(tmdb: TmdbClient, input: TmdbRecallIn
     } catch (e) {
       // 对照 MP get_info 失败返回 None 后继续:详情只用于补别名,取不到保留搜索结果
       if (!(e instanceof TmdbError)) throw e;
+      input.onDetailError?.(e, c);
     }
   }
   return [...seen.values()].slice(0, input.limit ?? 20);