nsfwapp/backend/analyze.go

// backend\analyze.go

package main

import (
	"bytes"
	"context"
	"encoding/json"
	"fmt"
	"image"
	"image/jpeg"
	"io"
	"math"
	"net/http"
	"os"
	"os/exec"
	"path/filepath"
	"runtime/debug"
	"sort"
	"strings"
	"sync"
	"time"
)

type analyzeVideoReq struct {
	JobID  string `json:"jobId"`
	Output string `json:"output"`
	Mode   string `json:"mode"` // "video"
	Goal   string `json:"goal"` // "highlights"
}

type analyzeHit struct {
	Time  float64 `json:"time"`
	Label string  `json:"label"`
	Score float64 `json:"score,omitempty"`
	Start float64 `json:"start,omitempty"`
	End   float64 `json:"end,omitempty"`
}

type analyzeVideoResp struct {
	OK       bool            `json:"ok"`
	Mode     string          `json:"mode,omitempty"`
	Goal     string          `json:"goal,omitempty"`
	Hits     []analyzeHit    `json:"hits"`
	Segments []aiSegmentMeta `json:"segments,omitempty"`
	Rating   *aiRatingMeta   `json:"rating,omitempty"`
	Error    string          `json:"error,omitempty"`
}

type videoFrameSample struct {
	Index int
	Time  float64
	Path  string
}

const (
	analyzeSegmentMergeGapSeconds = 8.0

	// Ein Label darf kurz verschwinden, ohne dass ein neues Segment entsteht.
	// Bei 3s Frame-Intervall heißt das: ein fehlender Frame wird überbrückt.
	analyzeLabelInvisibleGraceSeconds = 3.0

	// Video-Modus: extrahiert 1 Frame alle N Sekunden.
	// 1 = jeder Sekunde, 3 = alle 3 Sekunden, 5 = alle 5 Sekunden.
	analyzeVideoFrameIntervalSeconds = 1

	// AI-Server nicht mit tausenden Pfaden auf einmal fluten.
	analyzeFramePredictBatchSize = 32

	// > 0 = Frames auf diese Breite skalieren.
	// <= 0 = Originalgröße des Videos behalten.
	analyzeVideoFrameWidth = 0

	// Lokaler optionaler Python-Inference-Server.
	// Kann per Environment überschrieben werden:
	// AI_SERVER_URL=http://127.0.0.1:8765
	analyzeAIServerDefaultURL = "http://127.0.0.1:8765"
)

const (
	analyzeMinPositionScore = 0.30
	analyzeMinBodyScore     = 0.25
	analyzeMinObjectScore   = 0.28
	analyzeMinClothingScore = 0.40
	analyzeMinDetectorScore = 0.35

	analyzeMinSingleSeverity = 0.60
	analyzeMinSingleScore    = 0.25
	analyzeMinSingleQuality  = 0.24

	analyzeMinComboScore = 0.36
)

const (
	analyzePositionClipWindowSeconds = 3.0
	analyzePositionClipMinScore      = 0.22
	analyzePositionClipMinFrames     = 2
	analyzePositionConflictMargin    = 0.08
	analyzePositionMinStableSeconds  = 8.0
	analyzePositionShortMaxScore     = 0.62
)

type analyzePositionEvidence struct {
	Time        float64
	Start       float64
	End         float64
	Label       string
	Score       float64
	Source      string
	PersonCount int
	HasPose     bool
	HasContext  bool
	HasClip     bool
}

func isAnalyzeContextOnlyPositionLabel(label string) bool {
	return false
}

func isAnalyzeTimelinePositionLabel(label string) bool {
	label = strings.ToLower(strings.TrimSpace(label))
	label = strings.TrimPrefix(label, "position:")
	return !isNoSexPositionLabel(label) &&
		isKnownPositionLabel(label) &&
		!isAnalyzeContextOnlyPositionLabel(label)
}

func analyzeVideoFrameFilter(intervalSeconds int) string {
	if intervalSeconds <= 0 {
		intervalSeconds = 1
	}

	fps := fmt.Sprintf("fps=1/%d", intervalSeconds)

	// Originalgröße behalten.
	if analyzeVideoFrameWidth <= 0 {
		return fps
	}

	return fmt.Sprintf(
		"%s,scale=%d:-2:flags=fast_bilinear",
		fps,
		analyzeVideoFrameWidth,
	)
}

func analyzeSingleFrameFilter() string {
	// Originalgröße behalten.
	if analyzeVideoFrameWidth <= 0 {
		return ""
	}

	return fmt.Sprintf(
		"scale=%d:-2:flags=fast_bilinear",
		analyzeVideoFrameWidth,
	)
}

func autoSelectedAILabelSet() map[string]struct{} {
	grouped, err := trainingGroupedLabels()
	if err != nil {
		appLogln("⚠️ analyze labels fallback:", err)
		return map[string]struct{}{}
	}

	out := map[string]struct{}{}

	add := func(values []string) {
		for _, value := range values {
			label := strings.ToLower(strings.TrimSpace(value))
			if isNoSexPositionLabel(label) {
				continue
			}
			out[label] = struct{}{}
		}
	}

	add(grouped.BodyParts)
	add(grouped.Objects)
	add(grouped.Clothing)
	add(grouped.SexPositions)

	return out
}

var autoSelectedAILabelsOnce sync.Once
var autoSelectedAILabelsCache map[string]struct{}

func shouldAutoSelectAnalyzeHit(label string) bool {
	label = strings.ToLower(strings.TrimSpace(label))
	if isNoSexPositionLabel(label) {
		return false
	}

	if isIgnoredNSFWLabel(label) || isPersonSegmentLabel(label) {
		return false
	}

	labels := autoSelectedAILabelSet()
	if _, ok := labels[label]; ok {
		return true
	}

	// Robuster Fallback:
	// Auch wenn trainingGroupedLabels()/detection_labels.json gerade leer,
	// veraltet oder nicht ladbar ist, sollen bekannte Rating-/Analyze-Labels
	// trotzdem durchkommen.
	if bodyPartSeverityWeight(label) > 0 {
		return true
	}
	if objectSeverityWeight(label) > 0 {
		return true
	}
	if clothingSeverityWeight(label) > 0 {
		return true
	}
	if isKnownPositionLabel(label) && positionSeverityWeight(label) > 0 {
		return true
	}

	return false
}

var nsfwIgnoredLabels = map[string]struct{}{
	// Personen sollen nicht als interessante Segmente auftauchen.
	"person_male":   {},
	"person_female": {},
}

func isIgnoredNSFWLabel(label string) bool {
	label = strings.ToLower(strings.TrimSpace(label))
	_, ok := nsfwIgnoredLabels[label]
	return ok
}

func addHighlightResult(best map[string]float64, label string, score float64) {
	label = strings.ToLower(strings.TrimSpace(label))
	if isNoSexPositionLabel(label) {
		return
	}

	if score <= 0 {
		score = 1
	}

	if old, ok := best[label]; !ok || score > old {
		best[label] = score
	}
}

func addScoredHighlightLabels(best map[string]float64, prefix string, items []TrainingScoredLabel) {
	prefix = strings.ToLower(strings.TrimSpace(prefix))
	if prefix == "" {
		return
	}

	for _, item := range items {
		label := strings.ToLower(strings.TrimSpace(item.Label))
		if isNoSexPositionLabel(label) {
			continue
		}

		addHighlightResult(best, prefix+":"+label, item.Score)
	}
}

func trainingPredictionToHighlightResults(pred TrainingPrediction) []NsfwFrameResult {
	best := map[string]float64{}

	sexPosition := strings.ToLower(strings.TrimSpace(pred.SexPosition))
	if isAnalyzeTimelinePositionLabel(sexPosition) {
		addHighlightResult(best, "position:"+sexPosition, pred.SexPositionScore)
	}

	addScoredHighlightLabels(best, "body", pred.BodyPartsPresent)
	addScoredHighlightLabels(best, "object", pred.ObjectsPresent)
	addScoredHighlightLabels(best, "clothing", pred.ClothingPresent)

	for _, box := range pred.Boxes {
		label := strings.ToLower(strings.TrimSpace(box.Label))
		if isNoSexPositionLabel(label) {
			continue
		}

		if isIgnoredNSFWLabel(label) {
			continue
		}

		// Wichtig:
		// Keine beliebigen YOLO-/COCO-Labels als Highlights übernehmen.
		// Nur bewusst erlaubte Analyse-Labels anzeigen.
		if !shouldAutoSelectAnalyzeHit(label) {
			continue
		}

		addHighlightResult(best, "detector:"+label, box.Score)
	}

	// Kombis nur erzeugen, wenn wirklich Position + Zusatz vorhanden ist.
	if isAnalyzeTimelinePositionLabel(sexPosition) {
		positionScore := pred.SexPositionScore
		if positionScore <= 0 {
			positionScore = 1
		}

		addCombo := func(prefix string, items []TrainingScoredLabel) {
			for _, item := range items {
				label := strings.ToLower(strings.TrimSpace(item.Label))
				if isNoSexPositionLabel(label) {
					continue
				}

				score := item.Score
				if score <= 0 {
					score = 1
				}

				comboScore := math.Min(positionScore, score)
				addHighlightResult(best, "combo:position:"+sexPosition+"+"+prefix+":"+label, comboScore)
			}
		}

		addCombo("body", pred.BodyPartsPresent)
		addCombo("object", pred.ObjectsPresent)
		addCombo("clothing", pred.ClothingPresent)
	}

	out := make([]NsfwFrameResult, 0, len(best))

	for label, score := range best {
		out = append(out, NsfwFrameResult{
			Label: label,
			Score: score,
		})
	}

	sort.Slice(out, func(i, j int) bool {
		if out[i].Score == out[j].Score {
			return out[i].Label < out[j].Label
		}
		return out[i].Score > out[j].Score
	})

	return out
}

func pickHighlightResults(results []NsfwFrameResult) []NsfwFrameResult {
	out := make([]NsfwFrameResult, 0, len(results))

	for _, r := range results {
		label := strings.ToLower(strings.TrimSpace(r.Label))
		if isNoSexPositionLabel(label) {
			continue
		}

		score := r.Score
		if score <= 0 {
			score = 1
		}

		// Schwellen kannst du später pro Gruppe anders machen.
		switch {
		case strings.HasPrefix(label, "combo:"):
			if score < 0.35 {
				continue
			}
		case strings.HasPrefix(label, "position:"):
			if score < 0.30 {
				continue
			}
		case strings.HasPrefix(label, "object:"):
			if score < 0.30 {
				continue
			}
		case strings.HasPrefix(label, "clothing:"):
			if score < 0.30 {
				continue
			}
		case strings.HasPrefix(label, "body:"):
			if score < 0.30 {
				continue
			}
		case strings.HasPrefix(label, "detector:"):
			raw := strings.TrimPrefix(label, "detector:")
			if !shouldAutoSelectAnalyzeHit(raw) {
				continue
			}
			if score < 0.40 {
				continue
			}
		}

		out = append(out, NsfwFrameResult{
			Label: label,
			Score: score,
		})
	}

	sort.Slice(out, func(i, j int) bool {
		if out[i].Score == out[j].Score {
			return out[i].Label < out[j].Label
		}
		return out[i].Score > out[j].Score
	})

	return out
}

func predictFrameForAnalyze(ctx context.Context, img image.Image) TrainingPrediction {
	_ = ctx

	if !trainingRecognitionEnabled() {
		return trainingEmptyPrediction("recognition_disabled")
	}

	// Fallback für alte Sprite-/Image-Pfade:
	// Wenn kein echter Video-Frame-Pfad vorhanden ist, schreiben wir NICHT mehr in os.TempDir(),
	// sondern in generated/training/analyze-temp, damit die Bilder auffindbar bleiben.
	root, err := trainingRootDir()
	if err != nil {
		return trainingEmptyPrediction("root_failed")
	}

	tmpDir := filepath.Join(root, "analyze-temp")
	if err := os.MkdirAll(tmpDir, 0755); err != nil {
		return trainingEmptyPrediction("mkdir_failed")
	}

	tmp, err := os.CreateTemp(tmpDir, "training-highlight-frame-*.jpg")
	if err != nil {
		return trainingEmptyPrediction("temp_failed")
	}

	tmpPath := tmp.Name()

	// Wichtig: NICHT löschen, damit du die Bilder kontrollieren kannst.
	// defer os.Remove(tmpPath)

	if err := jpeg.Encode(tmp, img, &jpeg.Options{Quality: 92}); err != nil {
		_ = tmp.Close()
		return trainingEmptyPrediction("encode_failed")
	}

	if err := tmp.Close(); err != nil {
		return trainingEmptyPrediction("close_failed")
	}

	return trainingPredictFrame(tmpPath)
}

func predictFramePathForAnalyze(ctx context.Context, framePath string) TrainingPrediction {
	_ = ctx

	if !trainingRecognitionEnabled() {
		return trainingEmptyPrediction("recognition_disabled")
	}

	return trainingPredictFrame(framePath)
}

type analyzeBatchPredictReq struct {
	Paths        []string `json:"paths"`
	DetectorOnly bool     `json:"detectorOnly"`
	ImageSize    int      `json:"imageSize,omitempty"`
	Model        string   `json:"model,omitempty"`
}

type analyzeBatchPredictResp struct {
	OK          bool                 `json:"ok"`
	Predictions []TrainingPrediction `json:"predictions"`
	Error       string               `json:"error,omitempty"`
}

func analyzeAIServerURL() string {
	raw := strings.TrimSpace(os.Getenv("AI_SERVER_URL"))
	if raw == "" {
		raw = analyzeAIServerDefaultURL
	}

	return strings.TrimRight(raw, "/")
}

func trainingPredictFramePathsBatchForAnalyze(
	ctx context.Context,
	paths []string,
	detectorOnly bool,
) ([]TrainingPrediction, error) {
	cleanPaths := make([]string, 0, len(paths))

	for _, path := range paths {
		path = strings.TrimSpace(path)
		if path == "" {
			continue
		}

		cleanPaths = append(cleanPaths, path)
	}

	if len(cleanPaths) == 0 {
		return nil, appErrorf("keine frame-pfade für batch prediction")
	}

	if !trainingRecognitionEnabled() {
		out := make([]TrainingPrediction, 0, len(cleanPaths))
		for range cleanPaths {
			out = append(out, trainingEmptyPrediction("recognition_disabled"))
		}
		return out, nil
	}

	payload := analyzeBatchPredictReq{
		Paths:        cleanPaths,
		DetectorOnly: detectorOnly,
	}

	if analyzeVideoFrameWidth > 0 {
		payload.ImageSize = analyzeVideoFrameWidth
	}

	body, err := json.Marshal(payload)
	if err != nil {
		return nil, err
	}

	url := analyzeAIServerURL() + "/predict-batch"

	req, err := http.NewRequestWithContext(
		ctx,
		http.MethodPost,
		url,
		bytes.NewReader(body),
	)
	if err != nil {
		return nil, err
	}

	req.Header.Set("Content-Type", "application/json")
	addAIServerAuth(req)

	client := &http.Client{
		Timeout: 120 * time.Second,
	}

	res, err := client.Do(req)
	if err != nil {
		if ctxErr := ctx.Err(); ctxErr != nil {
			return nil, ctxErr
		}
		return nil, err
	}
	defer res.Body.Close()

	rawBody, readErr := io.ReadAll(res.Body)
	if readErr != nil {
		if ctxErr := ctx.Err(); ctxErr != nil {
			return nil, ctxErr
		}
		return nil, readErr
	}

	var parsed analyzeBatchPredictResp
	if err := json.Unmarshal(rawBody, &parsed); err != nil {
		if ctxErr := ctx.Err(); ctxErr != nil {
			return nil, ctxErr
		}

		appLogf(
			"❌ [analyze] AI server invalid JSON status=%d url=%s body=%s",
			res.StatusCode,
			url,
			strings.TrimSpace(string(rawBody)),
		)

		return nil, appErrorf(
			"AI server lieferte ungültiges JSON: HTTP %d: %s",
			res.StatusCode,
			strings.TrimSpace(string(rawBody)),
		)
	}

	if res.StatusCode < 200 || res.StatusCode >= 300 || !parsed.OK {
		msg := strings.TrimSpace(parsed.Error)
		if msg == "" {
			msg = fmt.Sprintf("AI server HTTP %d", res.StatusCode)
		}

		appLogf(
			"❌ [analyze] AI server error status=%d url=%s error=%s body=%s",
			res.StatusCode,
			url,
			msg,
			strings.TrimSpace(string(rawBody)),
		)

		return nil, appErrorf("%s", msg)
	}

	if len(parsed.Predictions) == 0 {
		return nil, appErrorf("AI server lieferte keine predictions")
	}

	return parsed.Predictions, nil
}

func extractVideoFrameAt(ctx context.Context, outPath string, atSec float64) (image.Image, error) {
	tmp, err := os.CreateTemp("", "nsfw-frame-*.jpg")
	if err != nil {
		return nil, err
	}
	tmpPath := tmp.Name()
	_ = tmp.Close()
	defer os.Remove(tmpPath)

	ffmpegPath := strings.TrimSpace(getSettings().FFmpegPath)
	if ffmpegPath == "" {
		ffmpegPath = "ffmpeg"
	}

	args := []string{
		"-ss", fmt.Sprintf("%.3f", atSec),
		"-i", outPath,
		"-frames:v", "1",
	}

	if vf := analyzeSingleFrameFilter(); vf != "" {
		args = append(args, "-vf", vf)
	}

	args = append(args,
		"-q:v", "2",
		"-y",
		tmpPath,
	)

	cmd := exec.CommandContext(ctx, ffmpegPath, args...)
	hideCommandWindow(cmd)

	if out, err := cmd.CombinedOutput(); err != nil {
		return nil, appErrorf("ffmpeg fehlgeschlagen: %v: %s", err, strings.TrimSpace(string(out)))
	}

	f, err := os.Open(tmpPath)
	if err != nil {
		return nil, err
	}
	defer f.Close()

	img, _, err := image.Decode(f)
	if err != nil {
		return nil, err
	}

	return img, nil
}

func analyzeFramesDirForOutput(outPath string) (string, error) {
	id := strings.TrimSpace(videoIDFromOutputPath(outPath))
	if id == "" {
		return "", appErrorf("konnte keine video-id aus output ableiten")
	}

	metaPath, err := generatedMetaFile(id)
	if err != nil || strings.TrimSpace(metaPath) == "" {
		return "", appErrorf("meta.json nicht gefunden")
	}

	return filepath.Join(filepath.Dir(metaPath), "frames"), nil
}

func cleanupAnalyzeFramesDirForOutput(outPath string) error {
	framesDir, err := analyzeFramesDirForOutput(outPath)
	if err != nil {
		return err
	}

	// Sicherheitscheck: niemals versehentlich etwas anderes löschen.
	if filepath.Base(framesDir) != "frames" {
		return appErrorf("cleanup abgebrochen: unerwarteter frames-ordner: %s", framesDir)
	}

	// Erst alle JPGs im frames-Ordner löschen.
	patterns := []string{
		filepath.Join(framesDir, "*.jpg"),
		filepath.Join(framesDir, "*.jpeg"),
	}

	for _, pattern := range patterns {
		files, globErr := filepath.Glob(pattern)
		if globErr != nil {
			return globErr
		}

		for _, file := range files {
			if removeErr := os.Remove(file); removeErr != nil && !os.IsNotExist(removeErr) {
				return removeErr
			}
		}
	}

	// Danach den Ordner selbst löschen.
	// Das klappt nur, wenn er leer ist. Falls dort andere Dateien liegen,
	// bleibt der Ordner absichtlich bestehen.
	if err := os.Remove(framesDir); err != nil && !os.IsNotExist(err) {
		return err
	}

	return nil
}

func extractVideoFramesBatch(
	ctx context.Context,
	outPath string,
	durationSec float64,
	intervalSeconds int,
	onExtracted func(current int, expected int),
) ([]videoFrameSample, func(), error) {
	if durationSec <= 0 {
		return nil, nil, appErrorf("videolänge fehlt")
	}

	if intervalSeconds <= 0 {
		intervalSeconds = 1
	}

	framesDir, err := analyzeFramesDirForOutput(outPath)
	if err != nil {
		return nil, nil, err
	}

	if err := os.MkdirAll(framesDir, 0755); err != nil {
		return nil, nil, err
	}

	// Alte Analyse-Frames entfernen, damit keine stale Frames mitgelesen werden.
	oldFrames, _ := filepath.Glob(filepath.Join(framesDir, "analyze-frame-*.jpg"))
	for _, oldFrame := range oldFrames {
		_ = os.Remove(oldFrame)
	}

	cleanup := func() {
		if err := cleanupAnalyzeFramesDirForOutput(outPath); err != nil {
			appLogln("⚠️ frames cleanup:", err)
		}
	}

	ffmpegPath := strings.TrimSpace(getSettings().FFmpegPath)
	if ffmpegPath == "" {
		ffmpegPath = "ffmpeg"
	}

	pattern := filepath.Join(framesDir, "analyze-frame-%06d.jpg")

	// fps=1/<intervalSeconds> bedeutet:
	// intervalSeconds=1  -> 1 Frame pro Sekunde
	// intervalSeconds=2  -> 1 Frame alle 2 Sekunden
	// intervalSeconds=5  -> 1 Frame alle 5 Sekunden
	vf := analyzeVideoFrameFilter(intervalSeconds)

	cmd := exec.CommandContext(
		ctx,
		ffmpegPath,
		"-hide_banner",
		"-loglevel", "error",
		"-i", outPath,
		"-vf", vf,
		"-q:v", "4",
		"-fps_mode", "vfr",
		"-y",
		pattern,
	)

	hideCommandWindow(cmd)

	expectedFrames := int(math.Ceil(durationSec / float64(intervalSeconds)))
	if expectedFrames < 1 {
		expectedFrames = 1
	}

	emitExtractProgress := func(current int) {
		if onExtracted == nil {
			return
		}

		if current < 0 {
			current = 0
		}
		if current > expectedFrames {
			current = expectedFrames
		}

		onExtracted(current, expectedFrames)
	}

	countExtractedFrames := func() int {
		files, err := filepath.Glob(filepath.Join(framesDir, "analyze-frame-*.jpg"))
		if err != nil {
			return 0
		}
		return len(files)
	}

	emitExtractProgress(0)

	var stderr bytes.Buffer
	cmd.Stderr = &stderr

	if err := cmd.Start(); err != nil {
		return nil, nil, appErrorf(
			"ffmpeg frames extrahieren konnte nicht gestartet werden: %w",
			err,
		)
	}

	done := make(chan error, 1)
	go func() {
		done <- cmd.Wait()
	}()

	lastCount := -1
	ticker := time.NewTicker(250 * time.Millisecond)
	defer ticker.Stop()

	waiting := true
	for waiting {
		select {
		case <-ctx.Done():
			if cmd.Process != nil {
				_ = cmd.Process.Kill()
			}

			select {
			case <-done:
			case <-time.After(2 * time.Second):
			}

			return nil, nil, ctx.Err()

		case waitErr := <-done:
			waiting = false

			count := countExtractedFrames()
			if count != lastCount {
				lastCount = count
				emitExtractProgress(count)
			}

			if waitErr != nil {
				msg := strings.TrimSpace(stderr.String())
				appLogf(
					"❌ [analyze] ffmpeg frame extract failed input=%q err=%v stderr=%s",
					outPath,
					waitErr,
					msg,
				)

				return nil, nil, appErrorf(
					"ffmpeg frames extrahieren fehlgeschlagen: %v: %s",
					waitErr,
					msg,
				)
			}

		case <-ticker.C:
			count := countExtractedFrames()
			if count != lastCount {
				lastCount = count
				emitExtractProgress(count)
			}
		}
	}

	files, err := filepath.Glob(filepath.Join(framesDir, "analyze-frame-*.jpg"))
	if err != nil {
		return nil, nil, err
	}

	sort.Strings(files)

	if len(files) == 0 {
		return nil, nil, appErrorf("ffmpeg hat keine frames erzeugt")
	}

	// Nach ffmpeg-Ende noch einmal den echten Endstand melden.
	// Das sorgt dafür, dass die Extraktionshälfte zuverlässig bei 50% landen kann.
	emitExtractProgress(len(files))

	out := make([]videoFrameSample, 0, len(files))

	for i, path := range files {
		t := float64(i * intervalSeconds)

		if t < 0 {
			t = 0
		}
		if durationSec > 0 && t > durationSec {
			t = durationSec
		}

		out = append(out, videoFrameSample{
			Index: i,
			Time:  t,
			Path:  path,
		})
	}

	return out, cleanup, nil
}

func recordAnalyzeVideo(w http.ResponseWriter, r *http.Request) {
	if !mustMethod(w, r, http.MethodPost) {
		return
	}

	var req analyzeVideoReq
	file := ""
	outPath := ""

	defer func() {
		if rec := recover(); rec != nil {
			msg := fmt.Sprintf("panic in analyse: %v", rec)
			appLogln("❌ [analyze]", msg)
			appLogln("❌ [analyze] stack:\n" + string(debug.Stack()))

			respondJSON(w, analyzeVideoResp{
				OK:    false,
				Mode:  "video",
				Goal:  "highlights",
				Hits:  []analyzeHit{},
				Error: msg,
			})
		}
	}()

	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
		logAnalyzeError("decode-request", "", "", err)
		respondJSON(w, analyzeVideoResp{
			OK:    false,
			Mode:  "video",
			Goal:  "highlights",
			Hits:  []analyzeHit{},
			Error: "ungültiger body: " + err.Error(),
		})
		return
	}

	req.Mode = "video"
	req.Goal = "highlights"

	outPath = strings.TrimSpace(req.Output)
	file = filepath.Base(outPath)

	appLogf("🧪 [analyze] request file=%q path=%q", file, outPath)

	if outPath == "" {
		err := appErrorf("output fehlt")
		logAnalyzeError("validate-output", file, outPath, err)

		respondJSON(w, analyzeVideoResp{
			OK:    false,
			Mode:  req.Mode,
			Goal:  req.Goal,
			Hits:  []analyzeHit{},
			Error: err.Error(),
		})
		return
	}

	fi, err := os.Stat(outPath)
	if err != nil || fi == nil || fi.IsDir() || fi.Size() <= 0 {
		if err == nil {
			err = appErrorf("output datei nicht gefunden oder leer")
		}
		logAnalyzeError("stat-output", file, outPath, err)

		respondJSON(w, analyzeVideoResp{
			OK:    false,
			Mode:  req.Mode,
			Goal:  req.Goal,
			Hits:  []analyzeHit{},
			Error: "output datei nicht gefunden: " + err.Error(),
		})
		return
	}

	appLogf(
		"🧪 [analyze] file ok file=%q size=%d mod=%s",
		file,
		fi.Size(),
		fi.ModTime().Format(time.RFC3339),
	)

	ctx, cancel := context.WithTimeout(r.Context(), 30*time.Minute)
	defer cancel()

	hits, analyzeStartedAtMs, _, err := analyzeVideoFromFrames(ctx, outPath)
	if err != nil {
		logAnalyzeError("analyze-frames", file, outPath, err)

		respondJSON(w, analyzeVideoResp{
			OK:    false,
			Mode:  req.Mode,
			Goal:  req.Goal,
			Hits:  []analyzeHit{},
			Error: err.Error(),
		})
		return
	}

	appLogf("🧪 [analyze] hits file=%q count=%d", file, len(hits))

	publishAnalyzePersistProgress(analyzeStartedAtMs, file, 0.1, "")

	durationSec, derr := durationSecondsForAnalyze(ctx, outPath)
	if derr != nil {
		logAnalyzeError("duration-after-analyze", file, outPath, derr)
	}
	if durationSec <= 0 {
		err := appErrorf("videolänge konnte nach analyse nicht bestimmt werden")
		logAnalyzeError("duration-invalid", file, outPath, err)
		publishAnalysisError(analyzeStartedAtMs, file, "Analyse fehlgeschlagen", err)

		respondJSON(w, analyzeVideoResp{
			OK:    false,
			Mode:  req.Mode,
			Goal:  req.Goal,
			Hits:  hits,
			Error: err.Error(),
		})
		return
	}

	segments := buildAnalyzeSegmentsForGoal(hits, durationSec)
	appLogf("🧪 [analyze] raw segments file=%q count=%d", file, len(segments))

	segments = prepareAIRatingSegments(segments)
	appLogf("🧪 [analyze] rating segments file=%q count=%d", file, len(segments))

	publishAnalyzePersistProgress(analyzeStartedAtMs, file, 0.45, "")

	rating := computeHighlightRatingForVideo(segments, durationSec, outPath)

	ai := &aiAnalysisMeta{
		Goal:           "highlights",
		Mode:           "video",
		Completed:      true,
		Hits:           hits,
		Segments:       segments,
		Rating:         rating,
		AnalyzedAtUnix: time.Now().Unix(),
	}

	persistCtx, persistCancel := context.WithTimeout(ctx, 90*time.Second)
	defer persistCancel()

	if err := writeVideoAIForFile(persistCtx, outPath, "", ai); err != nil {
		logAnalyzeError("write-meta-ai", file, outPath, err)
		publishAnalysisError(analyzeStartedAtMs, file, "Speichern fehlgeschlagen", err)

		respondJSON(w, analyzeVideoResp{
			OK:       false,
			Mode:     req.Mode,
			Goal:     req.Goal,
			Hits:     hits,
			Segments: segments,
			Rating:   rating,
			Error:    "analyse fertig, aber meta konnte nicht gespeichert werden: " + err.Error(),
		})
		return
	}

	autoDeleteLowRatedDownloadAfterAnalysis(persistCtx, outPath, rating)
	publishAnalyzePersistProgress(analyzeStartedAtMs, file, 1, "")
	publishAnalysisFinished(analyzeStartedAtMs, analyzeProgressTotal, file, "Analyse abgeschlossen")

	appLogf(
		"✅ [analyze] done file=%q hits=%d segments=%d rating=%.1f stars=%d",
		file,
		len(hits),
		len(segments),
		func() float64 {
			if rating == nil {
				return 0
			}
			return rating.Score
		}(),
		func() int {
			if rating == nil {
				return 0
			}
			return rating.Stars
		}(),
	)

	respondJSON(w, analyzeVideoResp{
		OK:       true,
		Mode:     req.Mode,
		Goal:     req.Goal,
		Hits:     hits,
		Segments: segments,
		Rating:   rating,
	})
}

type highlightSignal struct {
	Label string
	Score float64
	Group string
}

func normalizeHighlightSignalLabel(label string) string {
	label = strings.ToLower(strings.TrimSpace(label))
	if isNoSexPositionLabel(label) {
		return ""
	}

	switch {
	case strings.HasPrefix(label, "combo:"):
		// Bestehende alte Combos hier nicht weiterverwenden,
		// weil wir ab jetzt selbst saubere Kombis bauen.
		return ""

	case strings.HasPrefix(label, "detector:"):
		raw := strings.TrimPrefix(label, "detector:")
		if !shouldAutoSelectAnalyzeHit(raw) {
			return ""
		}
		return "detector:" + raw

	case strings.HasPrefix(label, "body:"):
		raw := strings.TrimPrefix(label, "body:")
		if isNoSexPositionLabel(raw) {
			return ""
		}
		return "body:" + raw

	case strings.HasPrefix(label, "object:"):
		raw := strings.TrimPrefix(label, "object:")
		if isNoSexPositionLabel(raw) {
			return ""
		}
		return "object:" + raw

	case strings.HasPrefix(label, "clothing:"):
		raw := strings.TrimPrefix(label, "clothing:")
		if isNoSexPositionLabel(raw) {
			return ""
		}
		return "clothing:" + raw

	case strings.HasPrefix(label, "position:"):
		raw := strings.TrimPrefix(label, "position:")
		if !isAnalyzeTimelinePositionLabel(raw) {
			return ""
		}
		return "position:" + raw

	default:
		if isIgnoredNSFWLabel(label) {
			return ""
		}

		if isAnalyzeTimelinePositionLabel(label) {
			return "position:" + label
		}

		if shouldAutoSelectAnalyzeHit(label) {
			return "detector:" + label
		}

		return ""
	}
}

func highlightSignalGroup(label string) string {
	label = strings.ToLower(strings.TrimSpace(label))

	switch {
	case strings.HasPrefix(label, "position:"):
		return "position"
	case strings.HasPrefix(label, "body:"):
		return "body"
	case strings.HasPrefix(label, "object:"):
		return "object"
	case strings.HasPrefix(label, "clothing:"):
		return "clothing"
	case strings.HasPrefix(label, "detector:"):
		raw := strings.TrimPrefix(label, "detector:")
		switch {
		case bodyPartSeverityWeight(raw) >= 0.65:
			return "body"
		case objectSeverityWeight(raw) >= 0.55:
			return "object"
		case clothingSeverityWeight(raw) >= 0.50:
			return "clothing"
		default:
			return "detector"
		}
	default:
		return "other"
	}
}

func rawAnalyzeSignalSeverityWeight(label string) float64 {
	label = normalizeHighlightSignalLabel(label)
	if label == "" {
		return 0
	}

	raw := normalizeSegmentLabel(label)
	if isNoSexPositionLabel(raw) {
		return 0
	}

	switch {
	case strings.HasPrefix(label, "position:"):
		return positionSeverityWeight(raw)

	case strings.HasPrefix(label, "body:"):
		return bodyPartSeverityWeight(raw)

	case strings.HasPrefix(label, "object:"):
		return objectSeverityWeight(raw)

	case strings.HasPrefix(label, "clothing:"):
		return clothingSeverityWeight(raw)

	case strings.HasPrefix(label, "detector:"):
		return detectorSeverityWeight(raw)

	default:
		return detectorSeverityWeight(raw)
	}
}

func highlightSignalInterestingEnough(label string, score float64) bool {
	label = normalizeHighlightSignalLabel(label)
	if label == "" {
		return false
	}

	if score <= 0 {
		score = 1
	}

	rawSev := rawAnalyzeSignalSeverityWeight(label)

	switch {
	case strings.HasPrefix(label, "position:"):
		return score >= analyzeMinPositionScore && rawSev > 0

	case strings.HasPrefix(label, "body:"):
		return score >= analyzeMinBodyScore && rawSev >= 0.65

	case strings.HasPrefix(label, "object:"):
		return score >= analyzeMinObjectScore && rawSev >= 0.50

	case strings.HasPrefix(label, "clothing:"):
		return score >= analyzeMinClothingScore && rawSev >= 0.50

	case strings.HasPrefix(label, "detector:"):
		return score >= analyzeMinDetectorScore && rawSev >= 0.60

	default:
		return false
	}
}

func addHighlightSignal(best map[string]highlightSignal, label string, score float64) {
	label = normalizeHighlightSignalLabel(label)
	if label == "" {
		return
	}

	if !highlightSignalInterestingEnough(label, score) {
		return
	}

	if score <= 0 {
		score = 1
	}

	key := normalizeSegmentLabel(label)
	if key == "" {
		return
	}

	sig := highlightSignal{
		Label: label,
		Score: score,
		Group: highlightSignalGroup(label),
	}

	if old, ok := best[key]; !ok || sig.Score > old.Score {
		best[key] = sig
	}
}

func addHighlightSignalsFromScoredLabels(
	best map[string]highlightSignal,
	prefix string,
	items []TrainingScoredLabel,
) {
	prefix = strings.ToLower(strings.TrimSpace(prefix))
	if prefix == "" {
		return
	}

	for _, item := range items {
		label := strings.ToLower(strings.TrimSpace(item.Label))
		if isNoSexPositionLabel(label) {
			continue
		}

		addHighlightSignal(best, prefix+":"+label, item.Score)
	}
}

func highlightComboPartOrder(label string) int {
	label = strings.ToLower(strings.TrimSpace(label))

	switch {
	case strings.HasPrefix(label, "position:"):
		return 0
	case strings.HasPrefix(label, "body:"):
		return 1
	case strings.HasPrefix(label, "object:"):
		return 2
	case strings.HasPrefix(label, "clothing:"):
		return 3
	case strings.HasPrefix(label, "detector:"):
		return 4
	default:
		return 9
	}
}

func predictionHasAnyAnalyzeSignal(pred TrainingPrediction) bool {
	sexPosition := strings.ToLower(strings.TrimSpace(pred.SexPosition))
	if !isNoSexPositionLabel(sexPosition) {
		return true
	}

	if len(pred.BodyPartsPresent) > 0 {
		return true
	}
	if len(pred.ObjectsPresent) > 0 {
		return true
	}
	if len(pred.ClothingPresent) > 0 {
		return true
	}
	if len(pred.Boxes) > 0 {
		return true
	}

	return false
}

func predictionUsableForAnalyze(pred TrainingPrediction) bool {
	if pred.ModelAvailable {
		return true
	}

	// Fallback:
	// Einige AI-Server/Antworten liefern verwertbare Labels/Boxes,
	// setzen modelAvailable aber nicht sauber.
	return predictionHasAnyAnalyzeSignal(pred)
}

func buildCombinedHighlightHitFromPrediction(pred TrainingPrediction, t float64) (analyzeHit, bool) {
	if !predictionUsableForAnalyze(pred) {
		return analyzeHit{}, false
	}

	best := map[string]highlightSignal{}

	sexPosition := strings.ToLower(strings.TrimSpace(pred.SexPosition))

	if isAnalyzeTimelinePositionLabel(sexPosition) {
		positionScore := pred.SexPositionScore
		if positionScore <= 0 {
			positionScore = 0.35
		}

		// Position soll als Kontext in Kombis bleiben.
		// Sie erzeugt weiterhin kein Segment alleine, weil unten mindestens
		// ein Nicht-Positionssignal verlangt wird.
		best["position:"+sexPosition] = highlightSignal{
			Label: "position:" + sexPosition,
			Score: math.Max(positionScore, 0.20),
			Group: "position",
		}
	}

	addHighlightSignalsFromScoredLabels(best, "body", pred.BodyPartsPresent)
	addHighlightSignalsFromScoredLabels(best, "object", pred.ObjectsPresent)
	addHighlightSignalsFromScoredLabels(best, "clothing", pred.ClothingPresent)

	for _, box := range pred.Boxes {
		label := strings.ToLower(strings.TrimSpace(box.Label))
		if isNoSexPositionLabel(label) {
			continue
		}
		if isIgnoredNSFWLabel(label) {
			continue
		}
		if !shouldAutoSelectAnalyzeHit(label) {
			continue
		}

		addHighlightSignal(best, "detector:"+label, box.Score)
	}

	signals := make([]highlightSignal, 0, len(best))
	groupSeen := map[string]bool{}
	nonPositionCount := 0
	hasPosition := false

	var bestSingle highlightSignal
	var bestSingleQuality float64

	for _, sig := range best {
		if sig.Label == "" {
			continue
		}

		if sig.Group == "position" {
			hasPosition = true
		} else {
			nonPositionCount++

			sev := segmentSeverityWeight(sig.Label)
			quality := sig.Score * sev

			if quality > bestSingleQuality {
				bestSingle = sig
				bestSingleQuality = quality
			}
		}

		groupSeen[sig.Group] = true
		signals = append(signals, sig)
	}

	// Fallback: starke Einzel-Treffer sollen auch Highlights werden.
	// Sonst verschwinden viele explizite Stellen, wenn sie nicht zufällig
	// im selben Frame mit Position/Object/Clothing kombiniert werden.
	returnSingleIfGoodEnough := func() (analyzeHit, bool) {
		if bestSingle.Label == "" {
			return analyzeHit{}, false
		}

		sev := segmentSeverityWeight(bestSingle.Label)

		// Nur wirklich relevante Einzel-Signale übernehmen.
		// Die Grenzwerte stehen oben als Konstanten, damit du sie später
		// leichter feintunen kannst:
		//
		// analyzeMinSingleSeverity = 0.60
		// analyzeMinSingleScore    = 0.25
		// analyzeMinSingleQuality  = 0.24
		if sev < analyzeMinSingleSeverity {
			return analyzeHit{}, false
		}

		if bestSingle.Score < analyzeMinSingleScore {
			return analyzeHit{}, false
		}

		if bestSingleQuality < analyzeMinSingleQuality {
			return analyzeHit{}, false
		}

		return analyzeHit{
			Time:  t,
			Label: bestSingle.Label,
			Score: bestSingle.Score,
			Start: t,
			End:   t,
		}, true
	}

	// Combo nur, wenn wirklich genug Kontext da ist.
	if len(signals) < 2 {
		return returnSingleIfGoodEnough()
	}
	if hasPosition && nonPositionCount < 1 {
		return returnSingleIfGoodEnough()
	}
	if !hasPosition && nonPositionCount < 2 {
		return returnSingleIfGoodEnough()
	}
	if len(groupSeen) < 2 && len(signals) < 3 {
		return returnSingleIfGoodEnough()
	}

	sort.SliceStable(signals, func(i, j int) bool {
		oi := highlightComboPartOrder(signals[i].Label)
		oj := highlightComboPartOrder(signals[j].Label)
		if oi != oj {
			return oi < oj
		}

		wi := segmentSeverityWeight(signals[i].Label) * signals[i].Score
		wj := segmentSeverityWeight(signals[j].Label) * signals[j].Score
		if wi != wj {
			return wi > wj
		}

		return signals[i].Label < signals[j].Label
	})

	// Nicht zu lange Titel bauen.
	if len(signals) > 4 {
		signals = signals[:4]
	}

	parts := make([]string, 0, len(signals))
	var scoreSum float64
	var scoreWeightSum float64
	var maxWeighted float64

	for _, sig := range signals {
		parts = append(parts, sig.Label)

		sev := segmentSeverityWeight(sig.Label)
		if sev <= 0 {
			sev = 0.5
		}

		weighted := sig.Score * sev
		scoreSum += sig.Score * sev
		scoreWeightSum += sev

		if weighted > maxWeighted {
			maxWeighted = weighted
		}
	}

	if len(parts) < 2 || scoreWeightSum <= 0 {
		return analyzeHit{}, false
	}

	avgScore := scoreSum / scoreWeightSum

	// Kombi-Score: stärkstes Signal + Durchschnitt.
	score := 0.65*maxWeighted + 0.35*avgScore
	if score > 1 {
		score = 1
	}
	if score <= 0 {
		score = avgScore
	}

	// Noch einmal Mindestqualität prüfen.
	if score < analyzeMinComboScore {
		return analyzeHit{}, false
	}

	return analyzeHit{
		Time:  t,
		Label: "combo:" + strings.Join(parts, "+"),
		Score: score,
		Start: t,
		End:   t,
	}, true
}

func buildHighlightHitsFromPrediction(pred TrainingPrediction, t float64) []analyzeHit {
	if !predictionUsableForAnalyze(pred) {
		return nil
	}

	best := map[string]highlightSignal{}

	sexPosition := strings.ToLower(strings.TrimSpace(pred.SexPosition))
	if isAnalyzeTimelinePositionLabel(sexPosition) {
		addHighlightSignal(best, "position:"+sexPosition, pred.SexPositionScore)
	}

	addHighlightSignalsFromScoredLabels(best, "body", pred.BodyPartsPresent)
	addHighlightSignalsFromScoredLabels(best, "object", pred.ObjectsPresent)
	addHighlightSignalsFromScoredLabels(best, "clothing", pred.ClothingPresent)

	for _, box := range pred.Boxes {
		label := strings.ToLower(strings.TrimSpace(box.Label))
		if isNoSexPositionLabel(label) {
			continue
		}
		if isIgnoredNSFWLabel(label) {
			continue
		}
		if !shouldAutoSelectAnalyzeHit(label) {
			continue
		}

		addHighlightSignal(best, "detector:"+label, box.Score)
	}

	out := make([]analyzeHit, 0, len(best))

	for _, sig := range best {
		label := strings.ToLower(strings.TrimSpace(sig.Label))
		if label == "" {
			continue
		}

		// Schwache Positions-Kontexte wie standing nicht alleine als Segment anzeigen.
		if sig.Group == "position" && segmentSeverityWeight(label) < 0.70 {
			continue
		}

		out = append(out, analyzeHit{
			Time:  t,
			Label: label,
			Score: sig.Score,
			Start: t,
			End:   t,
		})
	}

	sort.SliceStable(out, func(i, j int) bool {
		wi := segmentSeverityWeight(out[i].Label) * out[i].Score
		wj := segmentSeverityWeight(out[j].Label) * out[j].Score

		if wi != wj {
			return wi > wj
		}

		return out[i].Label < out[j].Label
	})

	return out
}

func appendHighlightHitsFromPrediction(
	hits []analyzeHit,
	pred TrainingPrediction,
	t float64,
) []analyzeHit {
	// Wichtig:
	// Wenn eine Kombi erkannt wurde, KEINE Einzel-Hits aus demselben Frame
	// zusätzlich anhängen. Sonst entstehen z. B. gleichzeitig:
	// - combo:position:toy_play+object:vibrator
	// - object:vibrator
	if combined, ok := buildCombinedHighlightHitFromPrediction(pred, t); ok {
		return append(hits, combined)
	}

	next := buildHighlightHitsFromPrediction(pred, t)
	if len(next) == 0 {
		return hits
	}

	return append(hits, next...)
}

func analyzeHitWithoutRawPosition(hit analyzeHit) (analyzeHit, bool) {
	label := strings.ToLower(strings.TrimSpace(hit.Label))
	if label == "" {
		return analyzeHit{}, false
	}

	if strings.HasPrefix(label, "position:") {
		return analyzeHit{}, false
	}

	if !strings.HasPrefix(label, "combo:") {
		hit.Label = label
		return hit, true
	}

	parts := []string{}
	for _, part := range strings.Split(strings.TrimPrefix(label, "combo:"), "+") {
		part = strings.ToLower(strings.TrimSpace(part))
		if part == "" || strings.HasPrefix(part, "position:") {
			continue
		}
		parts = append(parts, part)
	}

	switch len(parts) {
	case 0:
		return analyzeHit{}, false
	case 1:
		hit.Label = parts[0]
	default:
		hit.Label = "combo:" + strings.Join(parts, "+")
	}

	return hit, true
}

func appendVideoFrameHighlightHitsFromPrediction(
	hits []analyzeHit,
	pred TrainingPrediction,
	t float64,
) []analyzeHit {
	next := appendHighlightHitsFromPrediction(nil, pred, t)
	if len(next) == 0 {
		return hits
	}

	for _, hit := range next {
		if stripped, ok := analyzeHitWithoutRawPosition(hit); ok {
			hits = append(hits, stripped)
		}
	}

	return hits
}

func analyzePositionEvidenceFromPrediction(
	pred TrainingPrediction,
	t float64,
) (analyzePositionEvidence, bool) {
	if !predictionUsableForAnalyze(pred) {
		return analyzePositionEvidence{}, false
	}

	label := strings.ToLower(strings.TrimSpace(pred.SexPosition))
	if !isAnalyzeTimelinePositionLabel(label) {
		return analyzePositionEvidence{}, false
	}

	score := pred.SexPositionScore
	if score <= 0 {
		score = 0.35
	}
	score = clamp01(score)
	if score < 0.16 {
		return analyzePositionEvidence{}, false
	}

	source := strings.ToLower(strings.TrimSpace(pred.Source))
	personCount := len(pred.Persons)
	if personCount == 0 {
		for _, box := range pred.Boxes {
			if trainingIsPersonLikeLabel(box.Label) {
				personCount++
			}
		}
	}

	return analyzePositionEvidence{
		Time:        t,
		Start:       t,
		End:         t,
		Label:       label,
		Score:       score,
		Source:      source,
		PersonCount: personCount,
		HasPose:     strings.Contains(source, "yolo_pose") || len(pred.Persons) > 0,
		HasContext:  strings.Contains(source, "box_context") || len(pred.Boxes) > 0,
	}, true
}

func analyzePositionEvidenceWeight(item analyzePositionEvidence) float64 {
	weight := 1.0

	if item.HasClip && (item.HasPose || item.HasContext) {
		weight = 1.70
	} else if item.HasClip {
		weight = 1.55
	} else if item.HasPose && item.HasContext {
		weight = 1.05
	} else if item.HasPose {
		weight = 0.76
	} else if item.HasContext {
		weight = 0.82
	}

	if item.PersonCount >= 2 {
		weight += 0.08
	}

	return weight
}

type analyzePositionAggregate struct {
	Label        string
	WeightedSum  float64
	WeightSum    float64
	Count        int
	PoseCount    int
	ContextCount int
	ClipCount    int
	Start        float64
	End          float64
	Marker       float64
	BestScore    float64
}

type analyzePositionCandidate struct {
	Agg   *analyzePositionAggregate
	Score float64
}

type analyzePositionLedgerEntry struct {
	Start      float64
	End        float64
	Center     float64
	Candidates []analyzePositionCandidate
	Winner     analyzePositionCandidate
	Conflict   bool
}

func analyzePositionEvidenceSpan(item analyzePositionEvidence) (float64, float64) {
	start := item.Start
	end := item.End

	if start == 0 && end == 0 {
		start = item.Time
		end = item.Time
	}
	if start < 0 {
		start = item.Time
	}
	if end < 0 {
		end = item.Time
	}
	if start > end {
		start, end = end, start
	}

	return start, end
}

func analyzePositionAggregateScore(agg *analyzePositionAggregate, requiredFrames int) (float64, bool) {
	if agg == nil || agg.WeightSum <= 0 {
		return 0, false
	}

	minCount := requiredFrames
	if agg.ClipCount > 0 {
		// Ein VideoMAE-Clip steht bereits für ein Zeitfenster und muss nicht
		// zusätzlich noch mehrere Einzel-Frames derselben Position haben.
		minCount = 1
	} else if agg.PoseCount > 0 && agg.ContextCount == 0 {
		// Pose-only braucht etwas mehr zeitliche Stabilität.
		if minCount < 3 {
			minCount = 3
		}
	}

	if agg.Count < minCount {
		return 0, false
	}

	avg := clamp01(agg.WeightedSum / agg.WeightSum)
	stability := clamp01(float64(agg.Count) / math.Max(float64(requiredFrames), 3))
	sourceBonus := 0.0
	if agg.ClipCount > 0 && (agg.PoseCount > 0 || agg.ContextCount > 0) {
		sourceBonus = 0.07
	} else if agg.ClipCount > 0 {
		sourceBonus = 0.06
	} else if agg.PoseCount > 0 && agg.ContextCount > 0 {
		sourceBonus = 0.04
	}

	score := clamp01(avg*(0.86+0.14*stability) + sourceBonus)
	if agg.ClipCount == 0 && agg.PoseCount == 0 {
		score = math.Min(score, trainingPositionContextMaxScore)
	}
	if agg.ClipCount == 0 && agg.ContextCount == 0 {
		score = math.Min(score, trainingPoseStrongUnconfirmedMaxScore)
	}

	if score < analyzePositionClipMinScore {
		return 0, false
	}

	return score, true
}

func analyzePositionCandidateConflict(best, other analyzePositionCandidate) bool {
	if best.Agg == nil || other.Agg == nil {
		return false
	}
	if best.Agg.Label == other.Agg.Label {
		return false
	}

	// VideoMAE darf bei gleicher Groessenordnung als zeitliches Hauptsignal
	// gewinnen. Wenn aber zwei Clip-Positionen fast gleich stark sind, ist
	// das echter Konflikt und wird lieber nicht hart segmentiert.
	if best.Agg.ClipCount > 0 {
		if other.Agg.ClipCount > 0 {
			return other.Score >= best.Score-analyzePositionConflictMargin
		}
		return false
	}

	if other.Agg.ClipCount > 0 {
		return true
	}

	return other.Score >= best.Score-analyzePositionConflictMargin
}

func analyzePositionLedgerCenters(evidence []analyzePositionEvidence) []float64 {
	centers := make([]float64, 0, len(evidence)*3)

	for _, item := range evidence {
		start, end := analyzePositionEvidenceSpan(item)
		center := item.Time
		if center < start || center > end {
			center = (start + end) / 2
		}

		centers = append(centers, center)

		if item.HasClip && end > start {
			centers = append(centers, start, end)
		}
	}

	sort.Float64s(centers)

	out := centers[:0]
	for _, center := range centers {
		if len(out) == 0 || math.Abs(center-out[len(out)-1]) > 0.20 {
			out = append(out, center)
		}
	}

	return out
}

func analyzePositionCandidatesForWindow(
	evidence []analyzePositionEvidence,
	windowStart float64,
	windowEnd float64,
	requiredFrames int,
) []analyzePositionCandidate {
	byLabel := map[string]*analyzePositionAggregate{}

	for _, item := range evidence {
		itemStart, itemEnd := analyzePositionEvidenceSpan(item)
		if itemEnd < windowStart-0.001 || itemStart > windowEnd+0.001 {
			continue
		}

		label := strings.ToLower(strings.TrimSpace(item.Label))
		if !isAnalyzeTimelinePositionLabel(label) {
			continue
		}

		score := clamp01(item.Score)
		if score <= 0 {
			continue
		}

		agg := byLabel[label]
		if agg == nil {
			agg = &analyzePositionAggregate{
				Label:  label,
				Start:  itemStart,
				End:    itemEnd,
				Marker: item.Time,
			}
			byLabel[label] = agg
		}

		weight := analyzePositionEvidenceWeight(item)
		agg.WeightedSum += score * weight
		agg.WeightSum += weight
		agg.Count++
		if item.HasPose {
			agg.PoseCount++
		}
		if item.HasContext {
			agg.ContextCount++
		}
		if item.HasClip {
			agg.ClipCount++
		}
		if itemStart < agg.Start {
			agg.Start = itemStart
		}
		if itemEnd > agg.End {
			agg.End = itemEnd
		}
		if score > agg.BestScore {
			agg.BestScore = score
			agg.Marker = item.Time
		}
	}

	candidates := make([]analyzePositionCandidate, 0, len(byLabel))
	for _, agg := range byLabel {
		score, ok := analyzePositionAggregateScore(agg, requiredFrames)
		if ok {
			candidates = append(candidates, analyzePositionCandidate{
				Agg:   agg,
				Score: score,
			})
		}
	}

	sort.SliceStable(candidates, func(i, j int) bool {
		if candidates[i].Score != candidates[j].Score {
			return candidates[i].Score > candidates[j].Score
		}
		if candidates[i].Agg.ClipCount != candidates[j].Agg.ClipCount {
			return candidates[i].Agg.ClipCount > candidates[j].Agg.ClipCount
		}
		return candidates[i].Agg.Label < candidates[j].Agg.Label
	})

	return candidates
}

func stabilizeAnalyzePositionHits(in []analyzeHit, duration float64) []analyzeHit {
	if len(in) == 0 {
		return []analyzeHit{}
	}

	out := make([]analyzeHit, 0, len(in))
	for _, hit := range in {
		label := strings.ToLower(strings.TrimSpace(hit.Label))
		position := strings.TrimPrefix(label, "position:")
		if !isAnalyzeTimelinePositionLabel(position) {
			continue
		}

		start := hit.Start
		end := hit.End
		if end < start {
			start, end = end, start
		}
		if duration > 0 {
			start = math.Max(0, math.Min(duration, start))
			end = math.Max(0, math.Min(duration, end))
		}

		span := end - start
		if duration >= 60 &&
			span < analyzePositionMinStableSeconds &&
			hit.Score < analyzePositionShortMaxScore {
			continue
		}

		hit.Label = "position:" + position
		hit.Start = start
		hit.End = end
		out = append(out, hit)
	}

	return mergeAnalyzeHits(out)
}

func buildAnalyzePositionLedger(
	evidence []analyzePositionEvidence,
	duration float64,
	requiredFrames int,
) []analyzePositionLedgerEntry {
	if len(evidence) == 0 {
		return []analyzePositionLedgerEntry{}
	}

	halfWindow := analyzePositionClipWindowSeconds / 2
	if halfWindow <= 0 {
		halfWindow = math.Max(1, float64(analyzeVideoFrameIntervalSeconds))
	}

	centers := analyzePositionLedgerCenters(evidence)
	ledger := make([]analyzePositionLedgerEntry, 0, len(centers))

	for _, center := range centers {
		windowStart := math.Max(0, center-halfWindow)
		windowEnd := center + halfWindow
		if duration > 0 {
			windowEnd = math.Min(duration, windowEnd)
		}
		if windowEnd <= windowStart {
			windowEnd = windowStart + math.Max(1, float64(analyzeVideoFrameIntervalSeconds))
			if duration > 0 {
				windowEnd = math.Min(duration, windowEnd)
			}
		}

		entry := analyzePositionLedgerEntry{
			Start:  windowStart,
			End:    windowEnd,
			Center: center,
		}
		entry.Candidates = analyzePositionCandidatesForWindow(
			evidence,
			windowStart,
			windowEnd,
			requiredFrames,
		)

		if len(entry.Candidates) > 0 {
			if len(entry.Candidates) > 1 &&
				analyzePositionCandidateConflict(entry.Candidates[0], entry.Candidates[1]) {
				entry.Conflict = true
			} else {
				entry.Winner = entry.Candidates[0]
			}
		}

		ledger = append(ledger, entry)
	}

	return ledger
}

func buildClipPositionHitsFromEvidence(
	evidence []analyzePositionEvidence,
	duration float64,
) []analyzeHit {
	if len(evidence) == 0 {
		return []analyzeHit{}
	}

	sort.SliceStable(evidence, func(i, j int) bool {
		if evidence[i].Time != evidence[j].Time {
			return evidence[i].Time < evidence[j].Time
		}
		return evidence[i].Label < evidence[j].Label
	})

	requiredFrames := analyzePositionClipMinFrames
	if len(evidence) < requiredFrames {
		requiredFrames = len(evidence)
	}
	if requiredFrames < 1 {
		requiredFrames = 1
	}

	ledger := buildAnalyzePositionLedger(evidence, duration, requiredFrames)
	hits := make([]analyzeHit, 0, len(ledger))

	var cur analyzeHit
	hasCur := false
	flush := func() {
		if !hasCur {
			return
		}
		if cur.End <= cur.Start {
			cur.End = cur.Start + math.Max(1, float64(analyzeVideoFrameIntervalSeconds))
			if duration > 0 {
				cur.End = math.Min(duration, cur.End)
			}
		}
		cur.Time = (cur.Start + cur.End) / 2
		hits = append(hits, cur)
		hasCur = false
	}

	for _, entry := range ledger {
		if entry.Conflict || entry.Winner.Agg == nil {
			flush()
			continue
		}

		label := "position:" + entry.Winner.Agg.Label
		if hasCur &&
			cur.Label == label &&
			entry.Start <= cur.End+analyzeHitContinuationGapSeconds() {
			if entry.End > cur.End {
				cur.End = entry.End
			}
			if entry.Winner.Score > cur.Score {
				cur.Score = entry.Winner.Score
			}
			continue
		}

		flush()
		cur = analyzeHit{
			Time:  entry.Center,
			Label: label,
			Score: entry.Winner.Score,
			Start: entry.Start,
			End:   entry.End,
		}
		hasCur = true
	}
	flush()

	return stabilizeAnalyzePositionHits(mergeAnalyzeHits(hits), duration)
}

func analyzeVideoFromFrames(ctx context.Context, outPath string) ([]analyzeHit, int64, int, error) {
	return analyzeVideoFromFramesForGoal(ctx, outPath)
}

const (
	analyzeProgressTotal          = 1000
	analyzeProgressExtractEnd     = 450
	analyzeProgressInferenceStart = analyzeProgressExtractEnd
	analyzeProgressInferenceEnd   = 850
	analyzeProgressVideoMAEStart  = analyzeProgressInferenceEnd
	analyzeProgressVideoMAEEnd    = 950
	analyzeProgressPersistStart   = analyzeProgressVideoMAEEnd
	analyzeProgressPersistEnd     = 990
)

func logAnalyzeError(stage string, file string, outPath string, err error) {
	if err == nil {
		return
	}

	appLogf(
		"❌ [analyze] stage=%s file=%q path=%q err=%v",
		strings.TrimSpace(stage),
		strings.TrimSpace(file),
		strings.TrimSpace(outPath),
		err,
	)
}

func logAnalyzeInfo(stage string, file string, format string, args ...any) {
	prefix := fmt.Sprintf("🧪 [analyze] stage=%s file=%q ", strings.TrimSpace(stage), strings.TrimSpace(file))
	appLogf(prefix+format, args...)
}

func publishAnalyzeExtractProgress(
	startedAtMs int64,
	file string,
	progress float64,
	message string,
) {
	publishAnalyzePhaseProgress(
		startedAtMs,
		file,
		"Analyse",
		0,
		analyzeProgressExtractEnd,
		progress,
		message,
	)
}

func analyzePercentMessage(currentFrame int, totalFrames int) string {
	if totalFrames <= 0 {
		totalFrames = 1
	}

	ratio := float64(currentFrame) / float64(totalFrames)
	ratio = math.Max(0, math.Min(1, ratio))

	percent := int(math.Round(ratio * 100))
	if percent < 0 {
		percent = 0
	}
	if percent > 100 {
		percent = 100
	}

	return fmt.Sprintf("Analyse %d%%", percent)
}

func publishAnalyzeInferenceProgress(
	startedAtMs int64,
	file string,
	currentFrame int,
	totalFrames int,
	message string,
) {
	if totalFrames <= 0 {
		totalFrames = 1
	}

	ratio := float64(currentFrame) / float64(totalFrames)
	ratio = math.Max(0, math.Min(1, ratio))

	publishAnalyzePhaseProgress(
		startedAtMs,
		file,
		"Analyse",
		analyzeProgressInferenceStart,
		analyzeProgressInferenceEnd,
		ratio,
		message,
	)
}

func publishAnalyzeVideoMAEProgress(startedAtMs int64, file string, currentClip int, totalClips int) {
	if totalClips <= 0 {
		totalClips = 1
	}

	ratio := float64(currentClip) / float64(totalClips)
	publishAnalyzePhaseProgress(
		startedAtMs,
		file,
		"Analyse",
		analyzeProgressVideoMAEStart,
		analyzeProgressVideoMAEEnd,
		ratio,
		"",
	)
}

func publishAnalyzePersistProgress(startedAtMs int64, file string, progress float64, message string) {
	publishAnalyzePhaseProgress(
		startedAtMs,
		file,
		"Speichern",
		analyzeProgressPersistStart,
		analyzeProgressPersistEnd,
		progress,
		message,
	)
}

func publishAnalyzePhaseProgress(
	startedAtMs int64,
	file string,
	phase string,
	rangeStart int,
	rangeEnd int,
	progress float64,
	message string,
) {
	progress = math.Max(0, math.Min(1, progress))

	if rangeEnd < rangeStart {
		rangeEnd = rangeStart
	}

	current := rangeStart + int(math.Round(progress*float64(rangeEnd-rangeStart)))
	if current < 0 {
		current = 0
	}
	if current > analyzeProgressTotal {
		current = analyzeProgressTotal
	}

	message = strings.TrimSpace(message)
	if message == "" || strings.EqualFold(message, "Analyse") || strings.Contains(message, "Frames werden extrahiert") {
		phase = strings.TrimSpace(phase)
		if phase == "" {
			phase = "Analyse"
		}
		message = fmt.Sprintf("%s %d%%", phase, analyzeGlobalPercentFromCurrent(current, analyzeProgressTotal))
	}

	publishAnalysisStep(startedAtMs, current, analyzeProgressTotal, file, message)
}

func analyzeGlobalPercentFromCurrent(current int, total int) int {
	if total <= 0 {
		total = analyzeProgressTotal
	}

	ratio := float64(current) / float64(total)
	ratio = math.Max(0, math.Min(1, ratio))

	percent := int(math.Round(ratio * 100))
	if percent < 0 {
		return 0
	}
	if percent > 100 {
		return 100
	}
	return percent
}

func analyzeGlobalPercentMessageFromCurrent(current int, total int) string {
	return fmt.Sprintf(
		"Analyse %d%%",
		analyzeGlobalPercentFromCurrent(current, total),
	)
}

func analyzeVideoFromFramesForGoal(
	ctx context.Context,
	outPath string,
) (highlightHits []analyzeHit, startedAtMs int64, totalFrames int, err error) {
	file := filepath.Base(strings.TrimSpace(outPath))
	startedAtMs = publishAnalysisStarted(file, analyzeProgressTotal, "Analyse 0%")

	publishAnalyzeExtractProgress(
		startedAtMs,
		file,
		0,
		"Analyse 0%",
	)

	if err := ctx.Err(); err != nil {
		publishAnalysisError(startedAtMs, file, "Analyse abgebrochen", err)
		return nil, startedAtMs, 0, err
	}

	durationSec, _ := durationSecondsForAnalyze(ctx, outPath)
	if err := ctx.Err(); err != nil {
		publishAnalysisError(startedAtMs, file, "Analyse abgebrochen", err)
		return nil, startedAtMs, 0, err
	}

	if durationSec <= 0 {
		err := appErrorf("videolänge konnte nicht bestimmt werden")
		publishAnalysisError(startedAtMs, file, "Analyse fehlgeschlagen", err)
		return nil, startedAtMs, 0, err
	}

	publishPercentRange := func(lastPercent *int, nextPercent int, current int, total int, extractPhase bool) {
		if total <= 0 {
			total = 1
		}

		if nextPercent < 0 {
			nextPercent = 0
		}
		if nextPercent > 100 {
			nextPercent = 100
		}

		if nextPercent <= *lastPercent {
			return
		}

		for p := *lastPercent + 1; p <= nextPercent; p++ {
			if extractPhase {
				ratio := float64(p) / 50.0
				if ratio < 0 {
					ratio = 0
				}
				if ratio > 1 {
					ratio = 1
				}

				publishAnalyzeExtractProgress(
					startedAtMs,
					file,
					ratio,
					"",
				)
			} else {
				inferenceCurrent := current
				inferenceTotal := total

				if p >= 50 {
					inferenceTotal = 50
					inferenceCurrent = p - 50
				}

				publishAnalyzeInferenceProgress(
					startedAtMs,
					file,
					inferenceCurrent,
					inferenceTotal,
					"",
				)
			}
		}

		*lastPercent = nextPercent
	}

	failCancelled := func() ([]analyzeHit, int64, int, error) {
		err := ctx.Err()
		if err == nil {
			err = context.Canceled
		}

		publishAnalysisError(startedAtMs, file, "Analyse abgebrochen", err)
		return nil, startedAtMs, 0, err
	}

	lastExtractPercent := 0

	samples, cleanup, err := extractVideoFramesBatch(
		ctx,
		outPath,
		durationSec,
		analyzeVideoFrameIntervalSeconds,
		func(current int, expected int) {
			if expected <= 0 {
				expected = 1
			}

			if current < 0 {
				current = 0
			}
			if current > expected {
				current = expected
			}

			ratio := float64(current) / float64(expected)
			ratio = math.Max(0, math.Min(1, ratio))

			globalPercent := int(math.Round(ratio * 50))
			publishPercentRange(
				&lastExtractPercent,
				globalPercent,
				current,
				expected,
				true,
			)
		},
	)

	if cleanup != nil {
		defer cleanup()
	}

	if ctx.Err() != nil {
		return failCancelled()
	}

	if err != nil {
		publishAnalysisError(startedAtMs, file, "Frames konnten nicht extrahiert werden", err)
		return nil, startedAtMs, 0, err
	}

	if len(samples) == 0 {
		err := appErrorf("keine frame-samples vorhanden")
		publishAnalysisError(startedAtMs, file, "Keine Frames vorhanden", err)
		return nil, startedAtMs, 0, err
	}

	total := len(samples)

	if lastExtractPercent < 50 {
		publishPercentRange(
			&lastExtractPercent,
			50,
			total,
			total,
			true,
		)
	}

	if ctx.Err() != nil {
		return failCancelled()
	}

	paths := make([]string, 0, len(samples))
	for _, sample := range samples {
		if ctx.Err() != nil {
			return failCancelled()
		}

		paths = append(paths, sample.Path)
	}

	lastInferencePercent := 50

	publishAnalyzeInferenceProgress(
		startedAtMs,
		file,
		0,
		total,
		"",
	)

	if ctx.Err() != nil {
		return failCancelled()
	}

	batchOK := true
	detectorOnly := false
	positionEvidence := []analyzePositionEvidence{}

	for startIdx := 0; startIdx < len(samples); startIdx += analyzeFramePredictBatchSize {
		if ctx.Err() != nil {
			return failCancelled()
		}

		endIdx := startIdx + analyzeFramePredictBatchSize
		if endIdx > len(samples) {
			endIdx = len(samples)
		}

		predictions, batchErr := trainingPredictFramePathsBatchForAnalyze(
			ctx,
			paths[startIdx:endIdx],
			detectorOnly,
		)

		if ctx.Err() != nil {
			return failCancelled()
		}

		if batchErr != nil || len(predictions) < endIdx-startIdx {
			if batchErr != nil {
				appLogf(
					"❌ [analyze] batch failed file=%q batch=%d:%d err=%v",
					file,
					startIdx,
					endIdx,
					batchErr,
				)
			} else {
				appLogf(
					"❌ [analyze] batch returned too few predictions file=%q batch=%d:%d got=%d expected=%d",
					file,
					startIdx,
					endIdx,
					len(predictions),
					endIdx-startIdx,
				)
			}

			appLogln("⚠️ video batch analyse fehlgeschlagen, fallback auf einzelbild-analyse")

			batchOK = false
			highlightHits = nil
			lastInferencePercent = 50

			publishAnalyzeInferenceProgress(
				startedAtMs,
				file,
				0,
				total,
				"",
			)

			break
		}

		for i := 0; i < endIdx-startIdx; i++ {
			if ctx.Err() != nil {
				return failCancelled()
			}

			sample := samples[startIdx+i]
			pred := predictions[i]

			if !pred.ModelAvailable && predictionHasAnyAnalyzeSignal(pred) {
				appLogf(
					"⚠️ [analyze] prediction has signals but modelAvailable=false file=%q t=%.3f sex=%q body=%d objects=%d clothing=%d boxes=%d",
					file,
					sample.Time,
					strings.TrimSpace(pred.SexPosition),
					len(pred.BodyPartsPresent),
					len(pred.ObjectsPresent),
					len(pred.ClothingPresent),
					len(pred.Boxes),
				)
			}

			highlightHits = appendVideoFrameHighlightHitsFromPrediction(highlightHits, pred, sample.Time)
			if item, ok := analyzePositionEvidenceFromPrediction(pred, sample.Time); ok {
				positionEvidence = append(positionEvidence, item)
			}
		}

		globalPercent := 50 + int(math.Round((float64(endIdx)/float64(total))*50))
		if globalPercent > 100 {
			globalPercent = 100
		}

		publishPercentRange(
			&lastInferencePercent,
			globalPercent,
			endIdx,
			total,
			false,
		)
	}

	if batchOK {
		if ctx.Err() != nil {
			return failCancelled()
		}

		if lastInferencePercent < 100 {
			publishPercentRange(
				&lastInferencePercent,
				100,
				total,
				total,
				false,
			)
		}

		highlightHits, positionEvidence = applyVideoMAEPositionClipsForAnalyze(
			ctx,
			file,
			samples,
			durationSec,
			highlightHits,
			positionEvidence,
			func(current int, total int) {
				publishAnalyzeVideoMAEProgress(startedAtMs, file, current, total)
			},
		)
		if ctx.Err() != nil {
			return failCancelled()
		}

		highlightHits = append(
			highlightHits,
			buildClipPositionHitsFromEvidence(positionEvidence, durationSec)...,
		)

		cleanHighlightHits := mergeAnalyzeHits(highlightHits)

		return cleanHighlightHits, startedAtMs, total, nil
	}

	// Fallback: langsame Einzelbild-Analyse, nur wenn der AI-Server-Batch fehlschlägt.
	for i, sample := range samples {
		if ctx.Err() != nil {
			return failCancelled()
		}

		pred := predictFramePathForAnalyze(ctx, sample.Path)

		if ctx.Err() != nil {
			return failCancelled()
		}

		highlightHits = appendVideoFrameHighlightHitsFromPrediction(highlightHits, pred, sample.Time)
		if item, ok := analyzePositionEvidenceFromPrediction(pred, sample.Time); ok {
			positionEvidence = append(positionEvidence, item)
		}

		current := i + 1

		globalPercent := 50 + int(math.Round((float64(current)/float64(total))*50))
		if globalPercent > 100 {
			globalPercent = 100
		}

		publishPercentRange(
			&lastInferencePercent,
			globalPercent,
			current,
			total,
			false,
		)
	}

	if ctx.Err() != nil {
		return failCancelled()
	}

	if lastInferencePercent < 100 {
		publishPercentRange(
			&lastInferencePercent,
			100,
			total,
			total,
			false,
		)
	}

	highlightHits, positionEvidence = applyVideoMAEPositionClipsForAnalyze(
		ctx,
		file,
		samples,
		durationSec,
		highlightHits,
		positionEvidence,
		func(current int, total int) {
			publishAnalyzeVideoMAEProgress(startedAtMs, file, current, total)
		},
	)
	if ctx.Err() != nil {
		return failCancelled()
	}

	highlightHits = append(
		highlightHits,
		buildClipPositionHitsFromEvidence(positionEvidence, durationSec)...,
	)

	cleanHighlightHits := mergeAnalyzeHits(highlightHits)

	return cleanHighlightHits, startedAtMs, total, nil
}

func sameAnalyzeComboLabel(a, b string) bool {
	a = strings.ToLower(strings.TrimSpace(a))
	b = strings.ToLower(strings.TrimSpace(b))

	if !strings.HasPrefix(a, "combo:") || !strings.HasPrefix(b, "combo:") {
		return false
	}

	parse := func(label string) (position string, parts map[string]bool) {
		raw := strings.TrimPrefix(label, "combo:")
		parts = map[string]bool{}

		for _, part := range strings.Split(raw, "+") {
			part = strings.ToLower(strings.TrimSpace(part))
			if part == "" {
				continue
			}

			if strings.HasPrefix(part, "position:") {
				position = strings.TrimPrefix(part, "position:")
				continue
			}

			normalized := normalizeSegmentLabel(part)
			if normalized != "" {
				parts[normalized] = true
			}
		}

		return position, parts
	}

	posA, partsA := parse(a)
	posB, partsB := parse(b)

	// Unterschiedliche klare Hauptpositionen nicht zusammenführen.
	// Beispiel: doggy != missionary
	if posA != "" && posB != "" && posA != posB {
		return false
	}

	// Wenn beide keine gemeinsame Kontext-Komponente haben, nicht mergen.
	// Beispiel:
	// combo:position:doggy+object:dildo
	// combo:position:doggy+clothing:lingerie
	// => kein gemeinsames Nicht-Positionssignal, also getrennt lassen.
	for part := range partsA {
		if partsB[part] {
			return true
		}
	}

	return false
}

func sameAnalyzeSegmentLabel(a, b string) bool {
	a = strings.ToLower(strings.TrimSpace(a))
	b = strings.ToLower(strings.TrimSpace(b))

	if strings.HasPrefix(a, "combo:") || strings.HasPrefix(b, "combo:") {
		return sameAnalyzeComboLabel(a, b)
	}

	return normalizeSegmentLabel(a) == normalizeSegmentLabel(b)
}

func preferAnalyzeSegmentLabel(a, b string) string {
	a = strings.ToLower(strings.TrimSpace(a))
	b = strings.ToLower(strings.TrimSpace(b))

	if a == "" {
		return b
	}
	if b == "" {
		return a
	}

	if strings.HasPrefix(a, "combo:") && strings.HasPrefix(b, "combo:") {
		if strings.Contains(a, "position:") && !strings.Contains(b, "position:") {
			return a
		}
		if strings.Contains(b, "position:") && !strings.Contains(a, "position:") {
			return b
		}
		if len(b) > len(a) {
			return b
		}
		return a
	}

	// body: ist meist semantisch sauberer als detector:
	if strings.HasPrefix(a, "body:") && !strings.HasPrefix(b, "body:") {
		return a
	}
	if strings.HasPrefix(b, "body:") && !strings.HasPrefix(a, "body:") {
		return b
	}

	// object:/clothing:/position: ebenfalls sauberer als detector:
	preferredPrefix := func(s string) bool {
		return strings.HasPrefix(s, "object:") ||
			strings.HasPrefix(s, "clothing:") ||
			strings.HasPrefix(s, "position:") ||
			strings.HasPrefix(s, "combo:")
	}

	if preferredPrefix(a) && strings.HasPrefix(b, "detector:") {
		return a
	}
	if preferredPrefix(b) && strings.HasPrefix(a, "detector:") {
		return b
	}

	// Sonst kürzeres Label behalten, z. B. breasts statt detector:breasts.
	if len(b) < len(a) {
		return b
	}

	return a
}

func segmentLabelParts(label string) []string {
	label = strings.ToLower(strings.TrimSpace(label))
	if label == "" {
		return nil
	}

	if strings.HasPrefix(label, "combo:") {
		raw := strings.TrimPrefix(label, "combo:")
		parts := strings.Split(raw, "+")

		out := make([]string, 0, len(parts))
		for _, part := range parts {
			part = strings.ToLower(strings.TrimSpace(part))
			if part != "" {
				out = append(out, part)
			}
		}

		return out
	}

	return []string{label}
}

func segmentPositionFromAnalyzeLabel(label string) string {
	for _, part := range segmentLabelParts(label) {
		part = strings.TrimSpace(part)
		part = strings.TrimPrefix(part, "position:")

		if isAnalyzeTimelinePositionLabel(part) {
			return part
		}
	}

	return ""
}

func segmentTagsFromAnalyzeLabel(label string) []string {
	parts := segmentLabelParts(label)
	if len(parts) == 0 {
		return nil
	}

	out := make([]string, 0, len(parts))
	seen := map[string]bool{}

	for _, part := range parts {
		part = strings.ToLower(strings.TrimSpace(part))
		if part == "" {
			continue
		}

		// Person nicht als Segment-Tag speichern.
		if isPersonSegmentLabel(part) {
			continue
		}

		if strings.HasPrefix(part, "position:") {
			pos := strings.TrimPrefix(part, "position:")
			if !isAnalyzeTimelinePositionLabel(pos) {
				continue
			}

			tag := "position:" + pos
			if !seen[tag] {
				seen[tag] = true
				out = append(out, tag)
			}

			continue
		}

		if !seen[part] {
			seen[part] = true
			out = append(out, part)
		}
	}

	if len(out) == 0 {
		return nil
	}

	return out
}

func analyzeHitContinuationGapSeconds() float64 {
	// Treffer bei 0s und 6s sollen bei 3s Sampling noch zusammengehören:
	// 0s erkannt, 3s nicht erkannt, 6s wieder erkannt.
	return float64(analyzeVideoFrameIntervalSeconds) + analyzeLabelInvisibleGraceSeconds + 0.25
}

func mergeAnalyzeHits(in []analyzeHit) []analyzeHit {
	if len(in) == 0 {
		return []analyzeHit{}
	}

	maxGap := analyzeHitContinuationGapSeconds()

	byLabel := map[string][]analyzeHit{}

	for _, h := range in {
		label := strings.ToLower(strings.TrimSpace(h.Label))
		if isNoSexPositionLabel(label) {
			continue
		}
		if isIgnoredNSFWLabel(label) || isPersonSegmentLabel(label) {
			continue
		}

		start := h.Start
		end := h.End

		if start < 0 && end < 0 {
			start = h.Time
			end = h.Time
		} else {
			if start < 0 {
				start = h.Time
			}
			if end < 0 {
				end = h.Time
			}
		}

		if start > end {
			start, end = end, start
		}

		h.Label = label
		h.Start = start
		h.End = end

		key := normalizeSegmentLabel(label)
		if key == "" {
			continue
		}

		byLabel[key] = append(byLabel[key], h)
	}

	out := make([]analyzeHit, 0, len(in))

	for _, items := range byLabel {
		if len(items) == 0 {
			continue
		}

		sort.SliceStable(items, func(i, j int) bool {
			if items[i].Start != items[j].Start {
				return items[i].Start < items[j].Start
			}
			if items[i].End != items[j].End {
				return items[i].End < items[j].End
			}
			return items[i].Label < items[j].Label
		})

		cur := items[0]

		for i := 1; i < len(items); i++ {
			n := items[i]
			gap := n.Start - cur.End

			if gap >= -0.25 && gap <= maxGap {
				cur.Label = preferAnalyzeSegmentLabel(cur.Label, n.Label)

				if n.Start < cur.Start {
					cur.Start = n.Start
				}
				if n.End > cur.End {
					cur.End = n.End
				}
				if n.Score > cur.Score {
					cur.Score = n.Score
				}

				cur.Time = (cur.Start + cur.End) / 2
				continue
			}

			out = append(out, cur)
			cur = n
		}

		out = append(out, cur)
	}

	sort.SliceStable(out, func(i, j int) bool {
		if out[i].Start != out[j].Start {
			return out[i].Start < out[j].Start
		}
		if out[i].End != out[j].End {
			return out[i].End < out[j].End
		}
		return normalizeSegmentLabel(out[i].Label) < normalizeSegmentLabel(out[j].Label)
	})

	return out
}

func inferAnalyzePointSpanSeconds(hits []analyzeHit, duration float64) float64 {
	const fallback = 10.0

	if len(hits) < 2 {
		return fallback
	}

	times := make([]float64, 0, len(hits))

	for _, h := range hits {
		t := h.Time

		if t < 0 {
			if h.Start >= 0 {
				t = h.Start
			} else if h.End >= 0 {
				t = h.End
			}
		}

		if t < 0 {
			continue
		}

		if duration > 0 {
			t = math.Max(0, math.Min(t, duration))
		}

		times = append(times, t)
	}

	if len(times) < 2 {
		return fallback
	}

	sort.Float64s(times)

	gaps := make([]float64, 0, len(times)-1)
	prev := times[0]

	for _, t := range times[1:] {
		gap := t - prev
		if gap > 0.05 {
			gaps = append(gaps, gap)
			prev = t
		}
	}

	if len(gaps) == 0 {
		return fallback
	}

	sort.Float64s(gaps)

	median := gaps[len(gaps)/2]
	if len(gaps)%2 == 0 {
		median = (gaps[len(gaps)/2-1] + gaps[len(gaps)/2]) / 2
	}

	// Ein einzelner Frame repräsentiert ungefähr seinen Sample-Abstand,
	// aber wir deckeln, damit Sparse-Hits nicht riesig werden.
	span := median * 0.90

	if span < 6 {
		span = 6
	}
	if span > 12 {
		span = 12
	}

	return span
}

func expandAnalyzePointToSpan(t, span, duration float64) (float64, float64) {
	if span <= 0 {
		span = 3
	}

	if t < 0 {
		t = 0
	}

	if duration > 0 {
		t = math.Max(0, math.Min(t, duration))
	}

	half := span / 2

	start := t - half
	end := t + half

	if start < 0 {
		start = 0
	}
	if duration > 0 && end > duration {
		end = duration
	}

	if end <= start {
		if duration > 0 {
			end = math.Min(duration, start+math.Max(1, span))
			if end <= start {
				start = math.Max(0, end-math.Max(1, span))
			}
		} else {
			end = start + math.Max(1, span)
		}
	}

	return start, end
}

type analyzeLabelSegmentPoint struct {
	Label string
	Start float64
	End   float64
	Time  float64
	Score float64
}

func isAllowedAnalyzeSegmentLabel(label string) bool {
	label = strings.ToLower(strings.TrimSpace(label))
	if isNoSexPositionLabel(label) {
		return false
	}
	if isIgnoredNSFWLabel(label) || isPersonSegmentLabel(label) {
		return false
	}

	if strings.HasPrefix(label, "combo:") {
		for _, part := range segmentLabelParts(label) {
			if isAllowedAnalyzeSegmentLabel(part) {
				return true
			}
		}
		return false
	}

	raw := normalizeSegmentLabel(label)
	if isNoSexPositionLabel(raw) {
		return false
	}
	if strings.HasPrefix(label, "position:") && !isAnalyzeTimelinePositionLabel(label) {
		return false
	}
	if isAnalyzeContextOnlyPositionLabel(raw) {
		return false
	}

	return shouldAutoSelectAnalyzeHit(raw) || isAnalyzeTimelinePositionLabel(raw)
}

func buildLabelContinuitySegmentsFromAnalyzeHits(
	hits []analyzeHit,
	duration float64,
) []aiSegmentMeta {
	if len(hits) == 0 || duration <= 0 {
		return []aiSegmentMeta{}
	}

	sampleSpan := math.Max(1.0, float64(analyzeVideoFrameIntervalSeconds))
	halfSample := sampleSpan / 2.0
	maxGap := analyzeHitContinuationGapSeconds()

	byLabel := map[string][]analyzeLabelSegmentPoint{}

	for _, hit := range hits {
		label := strings.ToLower(strings.TrimSpace(hit.Label))
		if !isAllowedAnalyzeSegmentLabel(label) {
			continue
		}

		start := hit.Start
		end := hit.End

		if start < 0 && end < 0 {
			start = hit.Time
			end = hit.Time
		} else {
			if start < 0 {
				start = hit.Time
			}
			if end < 0 {
				end = hit.Time
			}
		}

		if start > end {
			start, end = end, start
		}

		if start < 0 {
			start = hit.Time
		}
		if end < 0 {
			end = start
		}

		start = math.Max(0, math.Min(start, duration))
		end = math.Max(0, math.Min(end, duration))

		key := normalizeSegmentLabel(label)
		if key == "" {
			continue
		}

		score := hit.Score
		if score <= 0 {
			score = 1
		}

		markerTime := hit.Time
		if markerTime < 0 {
			markerTime = (start + end) / 2
		}
		if duration > 0 {
			markerTime = math.Max(0, math.Min(markerTime, duration))
		}

		byLabel[key] = append(byLabel[key], analyzeLabelSegmentPoint{
			Label: label,
			Start: start,
			End:   end,
			Time:  markerTime,
			Score: score,
		})
	}

	out := make([]aiSegmentMeta, 0)

	for _, points := range byLabel {
		if len(points) == 0 {
			continue
		}

		sort.SliceStable(points, func(i, j int) bool {
			if points[i].Start != points[j].Start {
				return points[i].Start < points[j].Start
			}
			if points[i].End != points[j].End {
				return points[i].End < points[j].End
			}
			return points[i].Label < points[j].Label
		})

		curLabel := points[0].Label
		curStartMarker := points[0].Start
		curEndMarker := points[0].End
		curScoreSum := points[0].Score
		curScoreCount := 1
		curMarkerTime := points[0].Time
		curBestScore := points[0].Score

		for i := 1; i < len(points); i++ {
			n := points[i]

			gap := n.Start - curEndMarker

			if gap >= -0.25 && gap <= maxGap {
				curLabel = preferAnalyzeSegmentLabel(curLabel, n.Label)

				if n.Start < curStartMarker {
					curStartMarker = n.Start
				}
				if n.End > curEndMarker {
					curEndMarker = n.End
				}

				curScoreSum += n.Score
				curScoreCount++
				if n.Score > curBestScore {
					curBestScore = n.Score
					curMarkerTime = n.Time
				}
				continue
			}

			segment := makeLabelContinuitySegment(
				curLabel,
				curStartMarker,
				curEndMarker,
				curMarkerTime,
				curScoreSum,
				curScoreCount,
				halfSample,
				duration,
			)

			if segment.DurationSeconds > 0 {
				out = append(out, segment)
			}

			curLabel = n.Label
			curStartMarker = n.Start
			curEndMarker = n.End
			curScoreSum = n.Score
			curScoreCount = 1
			curMarkerTime = n.Time
			curBestScore = n.Score
		}

		segment := makeLabelContinuitySegment(
			curLabel,
			curStartMarker,
			curEndMarker,
			curMarkerTime,
			curScoreSum,
			curScoreCount,
			halfSample,
			duration,
		)

		if segment.DurationSeconds > 0 {
			out = append(out, segment)
		}
	}

	sort.SliceStable(out, func(i, j int) bool {
		if out[i].StartSeconds != out[j].StartSeconds {
			return out[i].StartSeconds < out[j].StartSeconds
		}
		if out[i].EndSeconds != out[j].EndSeconds {
			return out[i].EndSeconds < out[j].EndSeconds
		}
		return normalizeSegmentLabel(out[i].Label) < normalizeSegmentLabel(out[j].Label)
	})

	return out
}

func makeLabelContinuitySegment(
	label string,
	startMarker float64,
	endMarker float64,
	markerTime float64,
	scoreSum float64,
	scoreCount int,
	halfSample float64,
	duration float64,
) aiSegmentMeta {
	label = strings.ToLower(strings.TrimSpace(label))

	start := startMarker - halfSample
	end := endMarker + halfSample

	if start < 0 {
		start = 0
	}
	if duration > 0 && end > duration {
		end = duration
	}

	if end <= start {
		end = math.Min(duration, start+math.Max(1, halfSample*2))
	}

	score := 0.0
	if scoreCount > 0 {
		score = scoreSum / float64(scoreCount)
	}

	if markerTime < 0 {
		markerTime = (start + end) / 2
	}
	if duration > 0 {
		markerTime = math.Max(0, math.Min(markerTime, duration))
	}

	return aiSegmentMeta{
		Label:           label,
		StartSeconds:    start,
		EndSeconds:      end,
		DurationSeconds: math.Max(0, end-start),
		Score:           score,
		AutoSelected:    true,
		Position:        segmentPositionFromAnalyzeLabel(label),
		Tags:            segmentTagsFromAnalyzeLabel(label),
		MarkerSeconds:   markerTime,
		PreviewSeconds:  markerTime,
	}
}

func buildSegmentsFromAnalyzeHits(hits []analyzeHit, duration float64) []aiSegmentMeta {
	return buildLabelContinuitySegmentsFromAnalyzeHits(hits, duration)
}

func buildHighlightSegmentsFromAnalyzeHits(hits []analyzeHit, duration float64) []aiSegmentMeta {
	return buildLabelContinuitySegmentsFromAnalyzeHits(hits, duration)
}

func buildAnalyzeSegmentsForGoal(
	hits []analyzeHit,
	duration float64,
) []aiSegmentMeta {
	return buildHighlightSegmentsFromAnalyzeHits(hits, duration)
}

func durationSecondsForAnalyze(ctx context.Context, outPath string) (float64, error) {
	ctx2, cancel := context.WithTimeout(ctx, 8*time.Second)
	defer cancel()
	return durationSecondsCached(ctx2, outPath)
}

func videoIDFromOutputPath(outPath string) string {
	base := filepath.Base(strings.TrimSpace(outPath))
	if base == "" {
		return ""
	}
	stem := strings.TrimSuffix(base, filepath.Ext(base))
	stem = stripHotPrefix(stem)
	return strings.TrimSpace(stem)
}