fix(discovery): skip enrichment cache for date-less rows (year=0)

Rows without start_datum all hash to year=0, causing cache collisions
across unrelated markets. Gate both cache reads and writes on year!=0.

backend/internal/domain/discovery/service.go

@@ -717,20 +717,25 @@ func (s *Service) RunLLMEnrichOne(ctx context.Context, queueID uuid.UUID) (enric
 	if row.StartDatum != nil {
 		year = row.StartDatum.Year()
 	}
+	// Rows without a start_datum all hash to year=0, which would cause cache
+	// collisions across unrelated markets. Skip the cache entirely for these rows.
+	useCache := year != 0
 	cacheKey := enrich.CacheKey(row.NameNormalized, row.Stadt, year)
 
 	// Cache lookup — if we have a fresh LLM payload for this (name, city,
 	// year) tuple, skip the call. The merge still runs so a newly-populated
 	// crawl-enrich base gets its provenance preserved.
-	if cached, hit, err := s.repo.GetEnrichmentCache(ctx, cacheKey); err != nil {
-		slog.WarnContext(ctx, "enrichment cache get failed; continuing",
-			"cache_key", cacheKey, "error", err)
-	} else if hit {
-		merged := enrich.Merge(row.Enrichment, cached)
-		if err := s.repo.SetEnrichment(ctx, row.ID, merged, EnrichmentStatusDone); err != nil {
-			return enrich.Enrichment{}, fmt.Errorf("persist merged (cache hit): %w", err)
+	if useCache {
+		if cached, hit, err := s.repo.GetEnrichmentCache(ctx, cacheKey); err != nil {
+			slog.WarnContext(ctx, "enrichment cache get failed; continuing",
+				"cache_key", cacheKey, "error", err)
+		} else if hit {
+			merged := enrich.Merge(row.Enrichment, cached)
+			if err := s.repo.SetEnrichment(ctx, row.ID, merged, EnrichmentStatusDone); err != nil {
+				return enrich.Enrichment{}, fmt.Errorf("persist merged (cache hit): %w", err)
+			}
+			return merged, nil
 		}
-		return merged, nil
 	}
 
 	llmReq := enrich.LLMRequest{
@@ -753,10 +758,13 @@ func (s *Service) RunLLMEnrichOne(ctx context.Context, queueID uuid.UUID) (enric
 
 	// Cache the raw LLM output (not the merged result). A later re-crawl
 	// might change crawl-enrich fields; the cached answer should layer on
-	// top of whatever the current base is.
-	if err := s.repo.SetEnrichmentCache(ctx, cacheKey, llmPayload, enrich.DefaultCacheTTL); err != nil {
-		slog.WarnContext(ctx, "enrichment cache set failed; continuing",
-			"cache_key", cacheKey, "error", err)
+	// top of whatever the current base is. Skip for date-less rows (year=0)
+	// to avoid cross-row cache collisions.
+	if useCache {
+		if err := s.repo.SetEnrichmentCache(ctx, cacheKey, llmPayload, enrich.DefaultCacheTTL); err != nil {
+			slog.WarnContext(ctx, "enrichment cache set failed; continuing",
+				"cache_key", cacheKey, "error", err)
+		}
 	}
 
 	merged := enrich.Merge(row.Enrichment, llmPayload)

backend/internal/domain/discovery/service_test.go

@@ -853,6 +853,40 @@ func TestRunLLMEnrichOne_LLMErrorMarksFailed(t *testing.T) {
 	}
 }
 
+// TestRunLLMEnrichOne_SkipsCacheWhenYearZero: a row with no start_datum (year=0)
+// must bypass the enrichment cache entirely — both reads and writes. Two calls
+// for the same row must both hit the LLM.
+func TestRunLLMEnrichOne_SkipsCacheWhenYearZero(t *testing.T) {
+	rowID := uuid.New()
+	cacheHits := 0
+
+	repo := &mockRepo{
+		getDiscoveredFn: func(_ context.Context, _ uuid.UUID) (DiscoveredMarket, error) {
+			return DiscoveredMarket{
+				ID:             rowID,
+				MarktName:      "Testmarkt",
+				Stadt:          "Dresden",
+				NameNormalized: "testmarkt",
+				StartDatum:     nil, // no date → year = 0
+			}, nil
+		},
+		getCacheFn: func(_ string) (enrich.Enrichment, bool, error) {
+			cacheHits++
+			return enrich.Enrichment{}, false, nil
+		},
+	}
+	llm := &stubLLMEnricher{result: enrich.Enrichment{Category: catMittelaltermarkt}}
+	svc := NewService(repo, nil, noopLinkVerifier{}, noopMarketCreator{}, nil, llm, nil)
+
+	_, _ = svc.RunLLMEnrichOne(context.Background(), rowID)
+	_, _ = svc.RunLLMEnrichOne(context.Background(), rowID)
+
+	// With year=0, cache must be skipped — GetEnrichmentCache must never be called.
+	if cacheHits != 0 {
+		t.Errorf("expected 0 cache hits for date-less row, got %d", cacheHits)
+	}
+}
+
 // TestRunCrawlEnrichAll_EmptyQueueNoOp: nothing pending, zero summary, no writes.
 func TestRunCrawlEnrichAll_EmptyQueueNoOp(t *testing.T) {
 	var writes int