chore: remove Mistral/Ollama legacy references after Gemini migration

Rename mistral.go → llm_enricher.go and mistral_test.go →
llm_enricher_test.go; update all test function names and stale model
strings (mistral-large-latest → gemini-2.5-flash-lite); drop Ollama
block from .env; mark superseded planning specs; update provider
references in planning docs and CLAUDE.md to Google Gemini.

CLAUDE.md

@@ -26,7 +26,7 @@ Monorepo at `gitlab.com/vikingowl/marktvogt.de`. Components are regular director
 | Mobile         | Flutter                   |
 | Auth           | Custom (Go libs), E-Mail+PW / Magic Link / OAuth / 2FA |
 | Payments       | Stripe Connect            |
-| LLM            | OpenRouter                |
+| LLM            | Google Gemini             |
 | CI/CD          | GitLab CI (gitlab.com) — evaluation vs Woodpecker; sister project infinity-tales still on Woodpecker |
 | Hosting        | Kubernetes (itsh.dev)     |
 | Monitoring     | Prometheus, Loki, Grafana, Sentry |

backend/cmd/discovery-eval/README.md

@@ -3,10 +3,10 @@
 CLI that grades discovery's AI-backed components against labelled fixtures.
 Two modes:
 
-- `-mode similarity` (default) — `MistralSimilarityClassifier` on pair-
+- `-mode similarity` (default) — `SimilarityClassifier` on pair-
   labelled fixtures. Reports precision / recall / F1 / accuracy + a
   confidence calibration table.
-- `-mode category` — `MistralLLMEnricher`'s `category` output on row-
+- `-mode category` — `ProviderLLMEnricher`'s `category` output on row-
   labelled fixtures. Reports accuracy + a per-label confusion matrix.
 
 File-based cache keeps reruns free. Each mode has its own cache key shape,
@@ -18,7 +18,7 @@ so switching modes doesn't churn entries.
 
 ```
 export AI_API_KEY=...
-export AI_MODEL_COMPLEX=mistral-large-latest
+export AI_MODEL_COMPLEX=gemini-2.5-flash-lite
 
 go run ./backend/cmd/discovery-eval \
   -mode similarity \
@@ -34,7 +34,7 @@ Exit code is 1 when `F1 < threshold` (0 = gating disabled).
 
 ```
 export AI_API_KEY=...
-export AI_MODEL_COMPLEX=mistral-large-latest
+export AI_MODEL_COMPLEX=gemini-2.5-flash-lite
 
 go run ./backend/cmd/discovery-eval \
   -mode category \

backend/internal/domain/discovery/enrich/crawl_test.go

@@ -166,7 +166,7 @@ func TestMerge_BaseWinsOverOverlay(t *testing.T) {
 		Description:  "Ein großer Markt.",
 		InputTokens:  100,
 		OutputTokens: 50,
-		Model:        "mistral-large-latest",
+		Model:        "gemini-2.5-flash-lite",
 		Sources: Sources{
 			"plz":         ProvenanceLLM,
 			"venue":       ProvenanceLLM,

backend/internal/domain/discovery/enrich/llm_enricher_test.go

@@ -23,7 +23,7 @@ func (s *stubScraper) Fetch(_ context.Context, url string) (string, error) {
 	return s.responses[url], nil
 }
 
-func TestMistralEnrich_HappyPath(t *testing.T) {
+func TestLLMEnricher_HappyPath(t *testing.T) {
 	scraper := &stubScraper{responses: map[string]string{
 		"https://a.example/markt": "Ein Mittelaltermarkt mit Ritterspielen und Markttreiben.",
 		"https://b.example/info":  "Sa-So jeweils 10-18 Uhr.",
@@ -79,7 +79,7 @@ func TestMistralEnrich_HappyPath(t *testing.T) {
 	}
 }
 
-func TestMistralEnrich_AllScrapesFail(t *testing.T) {
+func TestLLMEnricher_AllScrapesFail(t *testing.T) {
 	scraper := &stubScraper{errs: map[string]error{
 		"https://a.example": errors.New("timeout"),
 		"https://b.example": errors.New("404"),
@@ -99,7 +99,7 @@ func TestMistralEnrich_AllScrapesFail(t *testing.T) {
 	}
 }
 
-func TestMistralEnrich_SomeScrapesFailStillCallsLLM(t *testing.T) {
+func TestLLMEnricher_SomeScrapesFailStillCallsLLM(t *testing.T) {
 	// One URL fails, one succeeds — the LLM should still run with partial
 	// context, not fail because of the one bad source.
 	scraper := &stubScraper{
@@ -124,7 +124,7 @@ func TestMistralEnrich_SomeScrapesFailStillCallsLLM(t *testing.T) {
 	}
 }
 
-func TestMistralEnrich_EmptyFieldsNoProvenance(t *testing.T) {
+func TestLLMEnricher_EmptyFieldsNoProvenance(t *testing.T) {
 	// LLM returns empty strings for fields it can't support. Those fields
 	// must NOT appear in Sources — an empty provenance is misleading.
 	scraper := &stubScraper{responses: map[string]string{"https://a.example": "Content."}}
@@ -148,7 +148,7 @@ func TestMistralEnrich_EmptyFieldsNoProvenance(t *testing.T) {
 	}
 }
 
-func TestMistralEnrich_CapsURLsAtFive(t *testing.T) {
+func TestLLMEnricher_CapsURLsAtFive(t *testing.T) {
 	// Supply 7 URLs; only the first 5 should be fetched.
 	urls := []string{"u1", "u2", "u3", "u4", "u5", "u6", "u7"}
 	responses := map[string]string{}

backend/internal/domain/discovery/enrich/similarity_test.go

@@ -31,7 +31,7 @@ func TestSimilarityPairKey_DifferentInputsDifferentKeys(t *testing.T) {
 	}
 }
 
-func TestMistralSimilarity_HappyPath(t *testing.T) {
+func TestSimilarityClassifier_HappyPath(t *testing.T) {
 	stub := &stubProvider{
 		content: `{"same_market":true,"confidence":0.82,"reason":"Gleicher Name, gleiche Stadt, gleiches Jahr."}`,
 	}
@@ -65,7 +65,7 @@ func TestMistralSimilarity_HappyPath(t *testing.T) {
 	}
 }
 
-func TestMistralSimilarity_ClampsConfidence(t *testing.T) {
+func TestSimilarityClassifier_ClampsConfidence(t *testing.T) {
 	tests := []struct {
 		name     string
 		raw      string
@@ -89,15 +89,15 @@ func TestMistralSimilarity_ClampsConfidence(t *testing.T) {
 	}
 }
 
-func TestMistralSimilarity_PropagatesPass2Error(t *testing.T) {
-	c := NewSimilarityClassifier(&stubProvider{err: errors.New("mistral down")})
+func TestSimilarityClassifier_PropagatesPass2Error(t *testing.T) {
+	c := NewSimilarityClassifier(&stubProvider{err: errors.New("provider down")})
 	_, err := c.Classify(context.Background(), SimilarityRow{}, SimilarityRow{})
 	if err == nil {
 		t.Fatal("expected error; got nil")
 	}
 }
 
-func TestMistralSimilarity_RejectsBadJSON(t *testing.T) {
+func TestSimilarityClassifier_RejectsBadJSON(t *testing.T) {
 	c := NewSimilarityClassifier(&stubProvider{content: "not json at all"})
 	_, err := c.Classify(context.Background(), SimilarityRow{}, SimilarityRow{})
 	if err == nil {

backend/internal/domain/discovery/service_test.go

@@ -696,7 +696,7 @@ func TestRunLLMEnrichOne_HappyPath(t *testing.T) {
 		Category:     catMittelaltermarkt,
 		Description:  "Ein großer Markt in der Altstadt.",
 		Sources:      enrich.Sources{"category": enrich.ProvenanceLLM, "description": enrich.ProvenanceLLM},
-		Model:        "mistral-large-latest",
+		Model:        "gemini-2.5-flash-lite",
 		InputTokens:  500,
 		OutputTokens: 80,
 	}
@@ -841,7 +841,7 @@ func TestRunLLMEnrichOne_LLMErrorMarksFailed(t *testing.T) {
 			return nil
 		},
 	}
-	llm := &stubLLMEnricher{err: errors.New("mistral down")}
+	llm := &stubLLMEnricher{err: errors.New("provider down")}
 	svc := NewService(repo, nil, noopLinkVerifier{}, noopMarketCreator{}, nil, llm, nil)
 
 	_, err := svc.RunLLMEnrichOne(context.Background(), rowID)
@@ -993,7 +993,7 @@ func TestClassifySimilarPair_LLMErrorPropagates(t *testing.T) {
 			return nil
 		},
 	}
-	sim := &stubSimilarityClassifier{err: errors.New("mistral 500")}
+	sim := &stubSimilarityClassifier{err: errors.New("provider 500")}
 	svc := NewService(repo, nil, noopLinkVerifier{}, noopMarketCreator{}, nil, nil, sim)
 
 	_, err := svc.ClassifySimilarPair(context.Background(), aID, bID)

planning/13-programm.md

@@ -37,7 +37,7 @@ Workflow:
 
 ## Entscheidungen
 
-- [x] **LLM-Provider**: OpenRouter (viele Modelle verfuegbar, einfach austauschbar)
+- [x] **LLM-Provider**: Google Gemini (gemini-2.5-flash-lite default)
 - [x] **Bild-Upload**: Ja, auch Bilder und Flyer-Scans werden per LLM geparsed
 - [x] **Push bei Aenderungen**: Ja, User mit Erinnerung fuer diesen Markt werden benachrichtigt
 - [x] **PDF-Export**: Ja, Programm als PDF exportierbar (fuer Aushang vor Ort)

planning/16-admin.md

@@ -45,7 +45,7 @@ Internes Dashboard fuer Plattform-Betreiber. Nicht oeffentlich zugaenglich.
   - Programm-Parsing (PDF/Bild)
   - Bild-Moderation
   - Natuerliche Sprache Suche (spaeter)
-- Provider: OpenRouter (Modell-Dropdown)
+- Provider: Google Gemini (gemini-2.5-flash-lite default)
 - Fallback-Modell konfigurierbar
 - Token-Verbrauch / Kosten-Uebersicht pro Bereich
 - API-Key-Verwaltung

planning/17-roadmap.md

@@ -38,7 +38,7 @@ Ziel: Veranstalter uebernehmen ihre Eintraege, erstellen eigene Maerkte.
 - Markt beanspruchen ("Das ist mein Markt")
 - Markt anlegen / bearbeiten
 - Kontaktdaten hinterlegen
-- Programm-Verwaltung (manuell + LLM-Parsing via OpenRouter)
+- Programm-Verwaltung (manuell + LLM-Parsing via Google Gemini)
 - Mitarbeiter-Sub-Rolle (granulare Berechtigungen)
 - Wunschliste (gesuchte Kategorien/Epochen)
 
@@ -125,7 +125,7 @@ Ziel: Interne Verwaltung und Moderation.
 - Admin-Rollen (Super-Admin, Moderator, Support)
 - Role-Assume (Impersonate)
 - Moderation (Bilder, Bewertungen, Chat)
-- LLM-Konfiguration (Modell pro Bereich, OpenRouter)
+- LLM-Konfiguration (Modell pro Bereich, Google Gemini)
 - Monitoring (Grafana-Einbettung, Plattform-Statistiken)
 - Audit Log
 - Admin-Benachrichtigungen (Push, E-Mail, Matrix Webhook)

planning/plans/2026-04-24-pluggable-ai-and-local-research.md

@@ -1,3 +1,5 @@
+> **SUPERSEDED 2026-04-25**: Replaced by Gemini-only migration (commit 3ddfd87). Kept for historical context. See `backend/internal/pkg/ai/gemini.go` for the implementation.
+
 # Pluggable AI Provider + Local Research Orchestrator — Implementation Plan
 
 > **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.

planning/specs/2026-04-24-pluggable-ai-and-local-research.md

@@ -1,3 +1,5 @@
+> **SUPERSEDED 2026-04-25**: Replaced by Gemini-only migration (commit 3ddfd87). Kept for historical context. See `backend/internal/pkg/ai/gemini.go` for the implementation.
+
 # Spec: Pluggable AI Provider + Local Research Orchestrator
 
 - **Date:** 2026-04-24