优化

internal/service/organizer_directory_metadata.go

@@ -55,7 +55,7 @@ func (o *OrganizerService) lookupOrganizeMetadata(ctx context.Context, src, sour
 				continue
 			}
 		}
-		match := o.scraper.lookup(ctx, lib, media, candidate, year)
+		match := o.scraper.lookup(ctx, lib, media, candidate, year, false)
 		if match != nil && strings.TrimSpace(match.Title) != "" {
 			if !organizeMetadataMatchTrusted(candidate, year, match) {
 				if cache != nil {

internal/service/recognition_words.go

@@ -5,8 +5,10 @@ import (
 	"encoding/json"
 	"fmt"
 	"net/http"
+	"regexp"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"go.uber.org/zap"
@@ -61,8 +63,9 @@ func NewRecognitionWordsService(log *zap.Logger, repo *repository.Container) *Re
 }
 
 func (s *RecognitionWordsService) Config(ctx context.Context) RecognitionWordsConfig {
-	cfg := recognitionWordsConfig(ctx, s.repo)
-	cfg.RuleCount = len(parseRecognitionWordRules(recognitionWordsCombinedText(cfg)))
+	cfg, rules := recognitionWordsRuntime(ctx, s.repo)
+	cfg = cloneRecognitionWordsConfig(cfg)
+	cfg.RuleCount = len(rules)
 	return cfg
 }
 
@@ -80,7 +83,11 @@ func (s *RecognitionWordsService) SaveConfig(ctx context.Context, cfg Recognitio
 	if err != nil {
 		return err
 	}
-	return s.repo.Setting.Set(ctx, RecognitionWordsSharedURLsKey, string(rawURLs))
+	if err := s.repo.Setting.Set(ctx, RecognitionWordsSharedURLsKey, string(rawURLs)); err != nil {
+		return err
+	}
+	invalidateRecognitionWordsRuntime(s.repo)
+	return nil
 }
 
 func (s *RecognitionWordsService) SyncShared(ctx context.Context) (RecognitionWordsConfig, error) {
@@ -104,6 +111,7 @@ func (s *RecognitionWordsService) SyncShared(ctx context.Context) (RecognitionWo
 	if err := s.repo.Setting.Set(ctx, RecognitionWordsSyncedAtKey, now); err != nil {
 		return cfg, err
 	}
+	invalidateRecognitionWordsRuntime(s.repo)
 	return s.Config(ctx), nil
 }
 
@@ -120,11 +128,10 @@ func (s *RecognitionWordsService) Test(ctx context.Context, input string) Recogn
 }
 
 func ApplyRecognitionWords(ctx context.Context, repo *repository.Container, raw string) string {
-	cfg := recognitionWordsConfig(ctx, repo)
-	if !cfg.Enabled {
+	cfg, rules := recognitionWordsRuntime(ctx, repo)
+	if !cfg.Enabled || len(rules) == 0 {
 		return raw
 	}
-	rules := parseRecognitionWordRules(recognitionWordsCombinedText(cfg))
 	return applyRecognitionWordRules(raw, rules)
 }
 
@@ -189,9 +196,69 @@ func normalizeRecognitionWordURLs(values []string) []string {
 type recognitionWordRule struct {
 	raw         string
 	block       string
+	blockRE     *regexp.Regexp
 	replaceFrom string
 	replaceTo   string
+	replaceRE   *regexp.Regexp
 	offsetLeft  string
 	offsetRight string
 	offsetExpr  string
+	offsetRE    *regexp.Regexp
+}
+
+// recognitionWordsRuntime caches the parsed rule set (with pre-compiled
+// regexes) per repository. Reading five settings rows and parsing/recompiling
+// the whole shared word list used to happen on every query candidate, which is
+// the scraper's hottest path. Writes through SaveConfig/SyncShared invalidate
+// the entry immediately; the TTL only bounds staleness for out-of-process edits.
+type recognitionWordsRuntimeEntry struct {
+	repo      *repository.Container
+	cfg       RecognitionWordsConfig
+	rules     []recognitionWordRule
+	expiresAt time.Time
+}
+
+const recognitionWordsRuntimeTTL = 30 * time.Second
+
+var recognitionWordsRuntimeCache struct {
+	sync.RWMutex
+	entry recognitionWordsRuntimeEntry
+	valid bool
+}
+
+func recognitionWordsRuntime(ctx context.Context, repo *repository.Container) (RecognitionWordsConfig, []recognitionWordRule) {
+	now := time.Now()
+	recognitionWordsRuntimeCache.RLock()
+	cached := recognitionWordsRuntimeCache.entry
+	hit := recognitionWordsRuntimeCache.valid && cached.repo == repo && now.Before(cached.expiresAt)
+	recognitionWordsRuntimeCache.RUnlock()
+	if hit {
+		return cached.cfg, cached.rules
+	}
+
+	cfg := recognitionWordsConfig(ctx, repo)
+	rules := parseRecognitionWordRules(recognitionWordsCombinedText(cfg))
+	recognitionWordsRuntimeCache.Lock()
+	recognitionWordsRuntimeCache.entry = recognitionWordsRuntimeEntry{
+		repo:      repo,
+		cfg:       cfg,
+		rules:     rules,
+		expiresAt: now.Add(recognitionWordsRuntimeTTL),
+	}
+	recognitionWordsRuntimeCache.valid = true
+	recognitionWordsRuntimeCache.Unlock()
+	return cfg, rules
+}
+
+func invalidateRecognitionWordsRuntime(repo *repository.Container) {
+	recognitionWordsRuntimeCache.Lock()
+	if recognitionWordsRuntimeCache.entry.repo == repo {
+		recognitionWordsRuntimeCache.valid = false
+	}
+	recognitionWordsRuntimeCache.Unlock()
+}
+
+func cloneRecognitionWordsConfig(cfg RecognitionWordsConfig) RecognitionWordsConfig {
+	cfg.SharedURLs = append([]string(nil), cfg.SharedURLs...)
+	return cfg
 }

internal/service/recognition_words_cache_test.go

@@ -0,0 +1,49 @@
+package service
+
+import (
+	"testing"
+
+	"go.uber.org/zap"
+)
+
+// The rule set is cached process-wide, so a config write must invalidate it
+// immediately; otherwise the admin would have to wait out the TTL before a
+// saved word list takes effect.
+func TestRecognitionWordsCacheInvalidatedBySaveConfig(t *testing.T) {
+	repos := newOrganizerTestRepo(t)
+	svc := NewRecognitionWordsService(zap.NewNop(), repos)
+
+	if err := svc.SaveConfig(t.Context(), RecognitionWordsConfig{
+		Enabled:   true,
+		LocalText: "BADWORD => 好标题",
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if got := ApplyRecognitionWords(t.Context(), repos, "BADWORD"); got != "好标题" {
+		t.Fatalf("first apply = %q, want 好标题", got)
+	}
+
+	if err := svc.SaveConfig(t.Context(), RecognitionWordsConfig{
+		Enabled:   true,
+		LocalText: "BADWORD => 新标题",
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if got := ApplyRecognitionWords(t.Context(), repos, "BADWORD"); got != "新标题" {
+		t.Fatalf("cached rules were not invalidated after SaveConfig: got %q, want 新标题", got)
+	}
+}
+
+func TestRecognitionWordsDisabledReturnsRawInput(t *testing.T) {
+	repos := newOrganizerTestRepo(t)
+	svc := NewRecognitionWordsService(zap.NewNop(), repos)
+	if err := svc.SaveConfig(t.Context(), RecognitionWordsConfig{
+		Enabled:   false,
+		LocalText: "BADWORD => 好标题",
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if got := ApplyRecognitionWords(t.Context(), repos, "BADWORD"); got != "BADWORD" {
+		t.Fatalf("disabled recognition words changed input: got %q", got)
+	}
+}

internal/service/recognition_words_rules.go

@@ -31,6 +31,11 @@ func parseRecognitionWordRule(line string) recognitionWordRule {
 		case strings.Contains(part, "<>") && strings.Contains(part, ">>"):
 			beforeAfter := strings.SplitN(part, ">>", 2)
 			bounds := strings.SplitN(beforeAfter[0], "<>", 2)
+			// A malformed rule without "<>" would otherwise index past the
+			// slice; word lists are fetched from the network, so stay defensive.
+			if len(bounds) < 2 {
+				continue
+			}
 			rule.offsetLeft = strings.TrimSpace(bounds[0])
 			rule.offsetRight = strings.TrimSpace(bounds[1])
 			rule.offsetExpr = strings.TrimSpace(beforeAfter[1])
@@ -38,56 +43,89 @@ func parseRecognitionWordRule(line string) recognitionWordRule {
 			rule.block = part
 		}
 	}
+	compileRecognitionWordRule(&rule)
 	return rule
 }
 
+var recognitionReplacementRE = regexp.MustCompile(`\\([0-9]+)`)
+
 func normalizeRecognitionReplacement(value string) string {
-	re := regexp.MustCompile(`\\([0-9]+)`)
-	return re.ReplaceAllString(value, "$$$1")
+	return recognitionReplacementRE.ReplaceAllString(value, "$$$1")
+}
+
+// compileRecognitionWordRule pre-compiles every pattern in a rule so the hot
+// clean-query path never recompiles regexes per candidate.
+func compileRecognitionWordRule(rule *recognitionWordRule) {
+	if rule == nil {
+		return
+	}
+	if rule.block != "" {
+		if re, err := regexp.Compile(rule.block); err == nil {
+			rule.blockRE = re
+		}
+	}
+	if rule.replaceFrom != "" {
+		if re, err := regexp.Compile(rule.replaceFrom); err == nil {
+			rule.replaceRE = re
+		}
+	}
+	if rule.offsetExpr != "" && (rule.offsetLeft != "" || rule.offsetRight != "") {
+		if re, err := compileRecognitionOffsetRE(rule.offsetLeft, rule.offsetRight); err == nil {
+			rule.offsetRE = re
+		}
+	}
+}
+
+func compileRecognitionOffsetRE(left, right string) (*regexp.Regexp, error) {
+	leftPattern := firstNonEmpty(left, `^`)
+	rightPattern := firstNonEmpty(right, `$`)
+	return regexp.Compile(`(?i)(` + leftPattern + `)(\d{1,5})(` + rightPattern + `)`)
 }
 
 func applyRecognitionWordRules(raw string, rules []recognitionWordRule) string {
 	out := strings.TrimSpace(raw)
 	for _, rule := range rules {
 		if rule.block != "" {
-			out = applyRecognitionBlock(out, rule.block)
+			out = applyRecognitionBlock(out, rule)
 		}
 		if rule.replaceFrom != "" {
-			out = applyRecognitionReplace(out, rule.replaceFrom, rule.replaceTo)
+			out = applyRecognitionReplace(out, rule)
 		}
 		if rule.offsetLeft != "" || rule.offsetRight != "" {
-			out = applyRecognitionOffset(out, rule.offsetLeft, rule.offsetRight, rule.offsetExpr)
+			out = applyRecognitionOffset(out, rule)
 		}
 	}
 	return strings.Join(strings.Fields(out), " ")
 }
 
-func applyRecognitionBlock(raw, block string) string {
-	if re, err := regexp.Compile(block); err == nil {
-		return re.ReplaceAllString(raw, " ")
+func applyRecognitionBlock(raw string, rule recognitionWordRule) string {
+	if rule.blockRE != nil {
+		return rule.blockRE.ReplaceAllString(raw, " ")
 	}
-	return strings.ReplaceAll(raw, block, " ")
+	return strings.ReplaceAll(raw, rule.block, " ")
 }
 
-func applyRecognitionReplace(raw, from, to string) string {
-	if re, err := regexp.Compile(from); err == nil {
-		return re.ReplaceAllString(raw, to)
+func applyRecognitionReplace(raw string, rule recognitionWordRule) string {
+	if rule.replaceRE != nil {
+		return rule.replaceRE.ReplaceAllString(raw, rule.replaceTo)
 	}
-	return strings.ReplaceAll(raw, from, to)
+	return strings.ReplaceAll(raw, rule.replaceFrom, rule.replaceTo)
 }
 
-func applyRecognitionOffset(raw, left, right, expr string) string {
-	if strings.TrimSpace(expr) == "" {
+func applyRecognitionOffset(raw string, rule recognitionWordRule) string {
+	if strings.TrimSpace(rule.offsetExpr) == "" {
 		return raw
 	}
-	leftPattern := firstNonEmpty(left, `^`)
-	rightPattern := firstNonEmpty(right, `$`)
-	re, err := regexp.Compile(`(?i)(` + leftPattern + `)(\d{1,5})(` + rightPattern + `)`)
-	if err != nil {
-		return raw
+	re := rule.offsetRE
+	if re == nil {
+		compiled, err := compileRecognitionOffsetRE(rule.offsetLeft, rule.offsetRight)
+		if err != nil {
+			return raw
+		}
+		re = compiled
 	}
 	return re.ReplaceAllStringFunc(raw, func(match string) string {
-		return applyRecognitionOffsetMatch(re, match, expr)
+		return applyRecognitionOffsetMatch(re, match, rule.offsetExpr)
 	})
 }
 

internal/service/scraper.go

@@ -63,12 +63,17 @@ func (s *ScraperService) EnrichOneWithOptions(ctx context.Context, m *model.Medi
 		}
 	}
 
+	// Same stale-negative problem as the library path: a single-item retry
+	// (queue task / manual rescrape) must get a fresh provider round-trip.
+	if options.RetryNoMatch && s != nil {
+		s.lookupCache.clearNegatives()
+	}
 	candidates := scrapeQueryCandidatesWithRecognition(ctx, s.repo, m, lib)
 	var query string
 	match := (*Match)(nil)
 	for _, candidate := range candidates {
 		query = candidate
-		candidateMatch := s.lookup(ctx, lib, m, candidate, year)
+		candidateMatch := s.lookup(ctx, lib, m, candidate, year, options.ForceRematch)
 		if candidateMatch == nil {
 			continue
 		}

internal/service/scraper_library.go

@@ -13,7 +13,12 @@ import (
 // lookup runs the provider chain after local NFO has been considered:
 // TMDb -> Douban -> Bangumi -> TheTVDB. Douban and Bangumi do not require API
 // keys; providers that are unavailable or return an error are skipped.
-func (s *ScraperService) lookup(ctx context.Context, lib *model.Library, media *model.Media, query string, year int) *Match {
+//
+// Results are cached per (kind, theatrical, query, year) because every
+// episode of a show produces the same candidate; bypassCache forces a fresh
+// provider round-trip for user-triggered rematches but still refreshes the
+// cache so later episodes reuse the corrected result.
+func (s *ScraperService) lookup(ctx context.Context, lib *model.Library, media *model.Media, query string, year int, bypassCache bool) *Match {
 	kind := ""
 	if lib != nil {
 		kind = lib.Type
@@ -28,6 +33,21 @@ func (s *ScraperService) lookup(ctx context.Context, lib *model.Library, media *
 	} else if (normalizeOrganizeMediaType(kind) != "movie" || explicitEpisode) && mediaIsEpisodic(media, lib) {
 		kind = "tv"
 	}
+
+	cacheKey := scrapeLookupCacheKey(kind, query, year, isTheatrical)
+	if !bypassCache {
+		if cached, ok := s.lookupCache.get(cacheKey); ok {
+			return cached
+		}
+	}
+	match := s.lookupUncached(ctx, kind, isTheatrical, query, year)
+	// Always refresh the cache, even on bypassed (forced) lookups, so a
+	// corrected rematch replaces the stale entry for later episodes.
+	s.lookupCache.set(cacheKey, match)
+	return match
+}
+
+func (s *ScraperService) lookupUncached(ctx context.Context, kind string, isTheatrical bool, query string, year int) *Match {
 	if s.tmdb != nil && s.tmdb.Enabled() {
 		if match := s.lookupAutomaticTMDb(ctx, kind, query, year); match != nil {
 			match.Provider = "tmdb"
@@ -107,6 +127,13 @@ func (s *ScraperService) EnrichLibraryDetailed(ctx context.Context, libraryID st
 
 func (s *ScraperService) EnrichLibraryDetailedWithOptions(ctx context.Context, libraryID string, options ScrapeOptions) (EnrichLibraryResult, error) {
 	result := EnrichLibraryResult{LibraryID: libraryID}
+	// A manual retry must not reuse stale negative cache entries from the
+	// previous run, otherwise "重新刮削" looks like it did nothing. Positives
+	// are kept so the rest of the season still dedupes; the first re-queried
+	// episode refreshes the entry for the rows behind it.
+	if options.RetryNoMatch && s != nil {
+		s.lookupCache.clearNegatives()
+	}
 	rows, err := s.scrapeCandidateRows(ctx, libraryID, options)
 	if err != nil {
 		return result, err

internal/service/scraper_lookup_cache.go

@@ -0,0 +1,126 @@
+package service
+
+import (
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+)
+
+// A TV library scrapes one media row per episode, and every row in the same
+// show computes the same query candidate (the series folder title). Without a
+// cache each episode re-issues the identical TMDb/Douban/Bangumi/TheTVDB
+// search, so a 100-episode season costs 100x the network calls it needs.
+//
+// The cache lives on the ScraperService instance (not a package global) so that
+// two services with different provider configuration never share results. TTL
+// bounds staleness for out-of-band edits; explicit invalidation is unnecessary
+// because the key includes the effective media kind, theatrical flag, query
+// and year. Negative entries use a shorter TTL so a failed query is retried
+// sooner without manual intervention.
+const (
+	scrapeLookupCacheTTL         = 10 * time.Minute
+	scrapeLookupNegativeCacheTTL = 2 * time.Minute
+	scrapeLookupCacheMaxItems    = 1024
+)
+
+type scrapeLookupCache struct {
+	mu      sync.Mutex
+	entries map[string]scrapeLookupCacheEntry
+}
+
+type scrapeLookupCacheEntry struct {
+	match     *Match
+	expiresAt time.Time
+}
+
+func newScrapeLookupCache() *scrapeLookupCache {
+	return &scrapeLookupCache{entries: map[string]scrapeLookupCacheEntry{}}
+}
+
+func scrapeLookupCacheKey(kind, query string, year int, isTheatrical bool) string {
+	theatrical := "0"
+	if isTheatrical {
+		theatrical = "1"
+	}
+	return strings.ToLower(strings.TrimSpace(kind)) + "|" +
+		theatrical + "|" +
+		strconv.Itoa(year) + "|" +
+		strings.ToLower(strings.TrimSpace(query))
+}
+
+// get reports whether the key is cached. A hit may carry a nil match, which
+// means the provider chain already ran and found nothing (negative cache).
+// Use clearNegative before a user-triggered retry so stale negative entries
+// do not suppress the fresh provider round-trip.
+func (c *scrapeLookupCache) get(key string) (*Match, bool) {
+	if c == nil || key == "" {
+		return nil, false
+	}
+	now := time.Now()
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	item, ok := c.entries[key]
+	if !ok {
+		return nil, false
+	}
+	if now.After(item.expiresAt) {
+		delete(c.entries, key)
+		return nil, false
+	}
+	return cloneMatch(item.match), true
+}
+
+func (c *scrapeLookupCache) set(key string, match *Match) {
+	if c == nil || key == "" {
+		return
+	}
+	now := time.Now()
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if len(c.entries) >= scrapeLookupCacheMaxItems {
+		for k, item := range c.entries {
+			if now.After(item.expiresAt) || len(c.entries) >= scrapeLookupCacheMaxItems {
+				delete(c.entries, k)
+			}
+			if len(c.entries) < scrapeLookupCacheMaxItems {
+				break
+			}
+		}
+	}
+	ttl := scrapeLookupCacheTTL
+	if match == nil {
+		ttl = scrapeLookupNegativeCacheTTL
+	}
+	c.entries[key] = scrapeLookupCacheEntry{match: cloneMatch(match), expiresAt: now.Add(ttl)}
+}
+
+// clearNegatives drops cached misses so a user-triggered retry gets a fresh
+// provider round-trip instead of reusing a stale negative entry.
+func (c *scrapeLookupCache) clearNegatives() {
+	if c == nil {
+		return
+	}
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	for k, item := range c.entries {
+		if item.match == nil {
+			delete(c.entries, k)
+		}
+	}
+}
+
+// cloneMatch returns an independent copy so callers that mutate the result
+// (localized title preference, local metadata merge, fanart artwork) cannot
+// corrupt the cached entry or leak state between media rows.
+func cloneMatch(m *Match) *Match {
+	if m == nil {
+		return nil
+	}
+	out := *m
+	out.Languages = append([]string(nil), m.Languages...)
+	out.Countries = append([]string(nil), m.Countries...)
+	out.Genres = append([]string(nil), m.Genres...)
+	out.Aliases = append([]string(nil), m.Aliases...)
+	return &out
+}

internal/service/scraper_lookup_cache_test.go

@@ -0,0 +1,165 @@
+package service
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"sync/atomic"
+	"testing"
+
+	"github.com/glebarez/sqlite"
+	"go.uber.org/zap"
+	"gorm.io/gorm"
+
+	"github.com/truewhile/MeBox/internal/config"
+	"github.com/truewhile/MeBox/internal/model"
+	"github.com/truewhile/MeBox/internal/repository"
+)
+
+// Every episode of a show produces the same query candidate (the series folder
+// title). Without the lookup cache each episode re-issues the identical search,
+// so a whole season costs one request per episode. This asserts the provider is
+// hit once and subsequent episodes reuse the cached match.
+func TestEnrichLibraryReusesLookupCacheAcrossEpisodes(t *testing.T) {
+	var searchCalls int32
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		switch r.URL.Path {
+		case "/search/tv":
+			if r.URL.Query().Get("query") != "折腰" {
+				_ = json.NewEncoder(w).Encode(map[string]any{"results": []any{}})
+				return
+			}
+			atomic.AddInt32(&searchCalls, 1)
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"results": []map[string]any{{
+					"id":             296753,
+					"name":           "折腰",
+					"overview":       "正确的剧集条目",
+					"poster_path":    "/zheyao.jpg",
+					"first_air_date": "2025-05-13",
+					"origin_country": []string{"CN"},
+				}},
+			})
+		default:
+			http.NotFound(w, r)
+		}
+	}))
+	defer upstream.Close()
+
+	db, err := gorm.Open(sqlite.Open("file::memory:?cache=shared"), &gorm.Config{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := db.AutoMigrate(&model.Library{}, &model.Series{}, &model.Media{}); err != nil {
+		t.Fatal(err)
+	}
+	repos := repository.New(db)
+	cfg := &config.Config{}
+	cfg.Secrets.TMDbAPIKey = "test-key"
+	cfg.Secrets.TMDbAPIProxy = upstream.URL
+	cfg.Secrets.TMDbImageProxy = upstream.URL + "/images"
+	log := zap.NewNop()
+	scraper := NewScraperService(cfg, log, repos, NewTMDbProvider(cfg, log, nil), nil, nil, nil, NewHub(log))
+
+	lib := model.Library{Name: "OpenList · 刮削缓存测试库", Path: "cloud://openlist/scrape-lookup-cache", Type: "tv", Enabled: true}
+	if err := repos.DB.Create(&lib).Error; err != nil {
+		t.Fatal(err)
+	}
+	const episodeCount = 4
+	for episode := 1; episode <= episodeCount; episode++ {
+		media := model.Media{
+			LibraryID:    lib.ID,
+			Title:        "折腰",
+			Path:         "cloud://openlist/scrape-lookup-cache/折腰 (2025)/Season 1/折腰.S01E0" + string(rune('0'+episode)) + ".mkv",
+			SeasonNum:    1,
+			EpisodeNum:   episode,
+			ScrapeStatus: "pending",
+		}
+		if err := repos.DB.Create(&media).Error; err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	result, err := scraper.EnrichLibraryDetailedWithOptions(t.Context(), lib.ID, ScrapeOptions{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Matched != episodeCount {
+		t.Fatalf("matched = %d, want %d", result.Matched, episodeCount)
+	}
+	if calls := atomic.LoadInt32(&searchCalls); calls != 1 {
+		t.Fatalf("tmdb /search/tv called %d times for %d episodes; want 1 (cache should dedupe identical queries)", calls, episodeCount)
+	}
+}
+
+// A cached match must be handed out as an independent copy: mutating one row's
+// result (localized title preference, local metadata merge) must not bleed into
+// the next row.
+func TestScrapeLookupCacheReturnsIndependentCopies(t *testing.T) {
+	cache := newScrapeLookupCache()
+	original := &Match{
+		Title:     "折腰",
+		TMDbID:    296753,
+		Genres:    []string{"剧情"},
+		Countries: []string{"CN"},
+		Aliases:   []string{"Zhe Yao"},
+	}
+	key := scrapeLookupCacheKey("tv", "折腰", 2025, false)
+	cache.set(key, original)
+
+	first, ok := cache.get(key)
+	if !ok || first == nil {
+		t.Fatal("expected a cache hit")
+	}
+	first.Title = "mutated"
+	first.Genres[0] = "mutated"
+	first.Aliases = append(first.Aliases, "extra")
+
+	second, ok := cache.get(key)
+	if !ok || second == nil {
+		t.Fatal("expected a second cache hit")
+	}
+	if second.Title != "折腰" || second.Genres[0] != "剧情" || len(second.Aliases) != 1 {
+		t.Fatalf("cached match was mutated through a returned copy: %+v", second)
+	}
+}
+
+// Negative results are cached too: a query that matched nothing must not be
+// re-issued for every remaining episode of the same show.
+func TestScrapeLookupCacheStoresNegativeResults(t *testing.T) {
+	cache := newScrapeLookupCache()
+	key := scrapeLookupCacheKey("tv", "no-such-show", 0, false)
+	cache.set(key, nil)
+	if _, ok := cache.get(key); !ok {
+		t.Fatal("negative result should be cached to avoid repeated provider calls")
+	}
+}
+
+// The theatrical flag is part of the key: a theatrical feature gets a tv
+// fallback lookup, so its result must not be shared with (or returned for)
+// the same query issued for a non-theatrical media row.
+func TestScrapeLookupCacheKeySeparatesTheatrical(t *testing.T) {
+	plain := scrapeLookupCacheKey("movie", "query", 2024, false)
+	theatrical := scrapeLookupCacheKey("movie", "query", 2024, true)
+	if plain == theatrical {
+		t.Fatalf("theatrical flag must be part of the cache key: %q", plain)
+	}
+}
+
+// A manual "retry no match" run clears stale negative entries so the provider
+// chain is actually re-queried instead of short-circuiting on the cached miss.
+func TestScrapeLookupCacheClearNegativesKeepsPositives(t *testing.T) {
+	cache := newScrapeLookupCache()
+	negKey := scrapeLookupCacheKey("tv", "missing", 0, false)
+	posKey := scrapeLookupCacheKey("tv", "found", 0, false)
+	cache.set(negKey, nil)
+	cache.set(posKey, &Match{Title: "found"})
+	cache.clearNegatives()
+	if _, ok := cache.get(negKey); ok {
+		t.Fatal("negative entry should be cleared before a manual retry")
+	}
+	if got, ok := cache.get(posKey); !ok || got == nil || got.Title != "found" {
+		t.Fatalf("positive entry should survive clearNegatives: %+v", got)
+	}
+}

internal/service/scraper_query_clean.go

@@ -24,11 +24,15 @@ var noiseTokens = []string{
 	"hkfree", "yify", "rarbg", "ettv", "fgt", "tgx", "ctrlhd", "ntb", "flux", "qhstudio",
 
 	// 流媒体平台 / 字幕组 / 国家版本（动漫常见）
-	"netflix", "nf", "amzn", "hulu", "disney", "max", "hbo",
+	// 注意：不要把同时是常见英文单词的标记放进来（如 max / web / judas），
+	// 否则 "Mad Max"、"Web Therapy" 这类正常标题会被误删。裸 "WEB" 发布标记
+	// 通常位于分辨率之后，由 releaseBoundary 截断规则处理；"WEB-DL" 则由
+	// multiWordNoise 单独匹配。
+	"netflix", "nf", "amzn", "hulu", "disney", "hbo",
 	"linetv", "ourtv", "iqiyi", "youku", "bilibili", "qiyi", "krj",
 	"atvp", "appletv", "apple-tv", "tx", "txweb",
 	"crunchyroll", "funimation", "anidb", "horriblesubs", "subsplease",
-	"erai-raws", "judas", "asw", "smcat", "leopard-raws", "ohys-raws", "colortv",
+	"erai-raws", "asw", "smcat", "leopard-raws", "ohys-raws", "colortv",
 	"mweb", "ubweb", "hhweb", "adweb", "chdweb", "kurosawa", "qhstudio",
 
 	// 中文字幕标记
@@ -55,6 +59,20 @@ var releaseBoundaryTokenSet = map[string]struct{}{
 	"x264": {}, "x265": {}, "h264": {}, "h265": {}, "h266": {}, "hevc": {}, "avc": {}, "av1": {}, "vvc": {},
 }
 
+// weakReleaseBoundaryTokenSet are release tags that double as plausible title
+// words. Before any release signal has been seen they are kept as part of the
+// title ("Mad Max", "Web Therapy"), so a title is never truncated to nothing;
+// after a real signal they behave like any other tag.
+var weakReleaseBoundaryTokenSet = map[string]struct{}{
+	"bd": {}, "dvd": {}, "web": {},
+}
+
+// releaseSignalToken marks the position of an extracted year. The year itself
+// is not a title token, but its presence still proves the following tokens are
+// release tags ("复仇者联盟4.2019.BD.1080p" must drop "BD"). A control rune is
+// used so it can never collide with a real filename token.
+const releaseSignalToken = "\u0001"
+
 var dynamicReleaseBoundaryTokenRE = regexp.MustCompile(`(?i)^(?:\d{3,4}p|\d{2,3}fps)$`)
 
 // bracketedTag matches "[anything]", "(anything)" or "{anything}" segments.
@@ -86,7 +104,9 @@ func CleanQuery(raw string) (title string, year int) {
 	if m := yearPattern.FindStringSubmatch(lower); len(m) >= 2 {
 		if v, err := strconv.Atoi(m[1]); err == nil {
 			year = v
-			lower = strings.ReplaceAll(lower, m[1], " ")
+			// Keep a positional marker: the year is not a title token, but its
+			// presence arms release-tag truncation for what follows.
+			lower = strings.ReplaceAll(lower, m[1], " "+releaseSignalToken+" ")
 		}
 	}
 
@@ -110,17 +130,35 @@ func CleanQuery(raw string) (title string, year int) {
 	}
 	// 拆分后丢掉过短（≤1）且全为 ASCII 数字 / 字母的"碎片"，避免
 	// 「2」「0」「v」之类残留干扰 TMDb 搜索。中文字符不算碎片。
+	//
+	// seenReleaseBoundary 只有在遇到分辨率/编码等强标记后才置位，置位后其余
+	// ASCII 词一律视为发布尾巴丢弃；releaseSignalled 则宽松得多，年份标记也会
+	// 置位，它只用来武装弱标记（bd/dvd/web）。这样 "Web Therapy" 不会被截空，
+	// 而 "2019.Avatar.1080p" 这种年份前置的标题也不会因为年份标记把后面的
+	// 真实标题词当成尾巴丢掉。
 	out := make([]string, 0, 8)
 	seenReleaseBoundary := false
+	releaseSignalled := false
 	for _, w := range strings.Fields(lower) {
+		if w == releaseSignalToken {
+			releaseSignalled = true
+			continue
+		}
 		if dynamicReleaseBoundaryTokenRE.MatchString(w) {
+			releaseSignalled = true
 			seenReleaseBoundary = true
 			continue
 		}
-		if _, ok := noiseTokenSet[w]; ok {
-			if _, boundary := releaseBoundaryTokenSet[w]; boundary {
+		if _, boundary := releaseBoundaryTokenSet[w]; boundary {
+			if _, weak := weakReleaseBoundaryTokenSet[w]; weak && !releaseSignalled {
+				// Ambiguous tag that is also a plausible title word; keep it
+				// until a real release signal proves we are past the title.
+			} else {
+				releaseSignalled = true
 				seenReleaseBoundary = true
+				continue
 			}
+		} else if _, ok := noiseTokenSet[w]; ok {
 			continue
 		}
 		if seenReleaseBoundary && isASCIIWord(w) {

internal/service/scraper_query_clean_regression_test.go

@@ -0,0 +1,77 @@
+package service
+
+import "testing"
+
+// CleanQuery must not delete words that are also ordinary English title words
+// just because a release group reused them as a tag. Regression guard: "max"
+// and "web" used to be unconditional noise/boundary tokens, which turned
+// "Mad Max" into "mad" and "Web Therapy" into an empty query.
+func TestCleanQueryKeepsCommonEnglishTitleWords(t *testing.T) {
+	cases := []struct {
+		in        string
+		wantTitle string
+		wantYear  int
+	}{
+		{"Mad.Max.1979.1080p.BluRay.mkv", "mad max", 1979},
+		{"Max.Payne.2008.1080p.WEB-DL.mkv", "max payne", 2008},
+		{"Web.Therapy.S01E01.1080p.WEB-DL.mkv", "web therapy", 0},
+		{"The.Web.2019.1080p.mkv", "the web", 2019},
+	}
+	for _, tc := range cases {
+		t.Run(tc.in, func(t *testing.T) {
+			gotTitle, gotYear := CleanQuery(tc.in)
+			if gotTitle != tc.wantTitle || gotYear != tc.wantYear {
+				t.Errorf("CleanQuery(%q) = (%q, %d), want (%q, %d)",
+					tc.in, gotTitle, gotYear, tc.wantTitle, tc.wantYear)
+			}
+		})
+	}
+}
+
+// A title consisting only of an ambiguous tag plus a release tail must not be
+// truncated to nothing; the tag stays so the query still has a chance.
+func TestCleanQueryDoesNotTruncateAmbiguousTagToNothing(t *testing.T) {
+	for _, in := range []string{"Web.1080p.WEB-DL.mkv", "BD.720p.mkv"} {
+		title, _ := CleanQuery(in)
+		if title == "" {
+			t.Errorf("CleanQuery(%q) returned an empty title", in)
+		}
+	}
+}
+
+// A year-prefixed filename must keep its title: the year arms weak-tag
+// truncation but must not discard the real title words that follow it.
+func TestCleanQueryKeepsTitleAfterYearPrefix(t *testing.T) {
+	cases := map[string]string{
+		"2019.Avatar.1080p.BluRay.mkv":     "avatar",
+		"2024.Dune.Part.Two.2160p.WEB.mkv": "dune part two",
+	}
+	for in, want := range cases {
+		t.Run(in, func(t *testing.T) {
+			got, year := CleanQuery(in)
+			if got != want {
+				t.Errorf("CleanQuery(%q) = (%q, %d), want title %q", in, got, year, want)
+			}
+		})
+	}
+}
+
+// Release-tag truncation must still fire once a real release signal (an
+// extracted year, resolution or codec) has been seen, so tags after it are
+// dropped as before.
+func TestCleanQueryStillDropsTagsAfterReleaseSignal(t *testing.T) {
+	cases := map[string]string{
+		"复仇者联盟4.2019.BD.1080p.mkv":               "复仇者联盟4",
+		"The.Matrix.1999.1080p.WEB-DL.H265.mp4":  "the matrix",
+		"Oppenheimer.2023.2160p.UHD.BluRay.mkv":  "oppenheimer",
+		"Interstellar.2014.4k.hdr.dts.atmos.mkv": "interstellar",
+	}
+	for in, want := range cases {
+		t.Run(in, func(t *testing.T) {
+			got, _ := CleanQuery(in)
+			if got != want {
+				t.Errorf("CleanQuery(%q) = %q, want %q", in, got, want)
+			}
+		})
+	}
+}

internal/service/scraper_service.go

@@ -26,6 +26,10 @@ type ScraperService struct {
 	cache   *RuntimeCacheService
 	images  *ImageProxy
 
+	// Caches provider-chain results keyed by media kind/query/year. Per-instance
+	// so services with different provider config never share results.
+	lookupCache *scrapeLookupCache
+
 	// Serializes final sidecar replacement. Windows cannot rename over an
 	// existing file, and concurrent scrapes can target the same sidecar.
 	artworkWriteMu sync.Mutex
@@ -50,6 +54,7 @@ func NewScraperService(
 	return &ScraperService{
 		cfg: cfg, log: log, repo: repo,
 		tmdb: tmdb, bangumi: bangumi, thetvdb: thetvdb, fanart: fanart, adult: adultProvider, hub: hub,
+		lookupCache: newScrapeLookupCache(),
 	}
 }