// backend\training.go

package main

import (
	"bufio"
	"context"
	"crypto/sha1"
	"encoding/hex"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"math"
	"math/rand"
	"net/http"
	"net/url"
	"os"
	"os/exec"
	"path/filepath"
	"sort"
	"strconv"
	"strings"
	"sync"
	"syscall"
	"time"
)

const trainingUncertainCandidateCount = 10

type TrainingLabels struct {
	People       []string `json:"people"`
	SexPositions []string `json:"sexPositions"`
	BodyParts    []string `json:"bodyParts"`
	Objects      []string `json:"objects"`
	Clothing     []string `json:"clothing"`
}

type TrainingBox struct {
	Label string  `json:"label"`
	Score float64 `json:"score,omitempty"`
	X     float64 `json:"x"`
	Y     float64 `json:"y"`
	W     float64 `json:"w"`
	H     float64 `json:"h"`
}

type TrainingScoredLabel struct {
	Label string  `json:"label"`
	Score float64 `json:"score"`
}

type TrainingPrediction struct {
	ModelAvailable bool   `json:"modelAvailable"`
	Source         string `json:"source,omitempty"`

	SexPosition      string                `json:"sexPosition"`
	SexPositionScore float64               `json:"sexPositionScore"`
	PeoplePresent    []TrainingScoredLabel `json:"peoplePresent"`
	BodyPartsPresent []TrainingScoredLabel `json:"bodyPartsPresent"`
	ObjectsPresent   []TrainingScoredLabel `json:"objectsPresent"`
	ClothingPresent  []TrainingScoredLabel `json:"clothingPresent"`
	Boxes            []TrainingBox         `json:"boxes"`
	Persons          []TrainingPosePerson  `json:"persons,omitempty"`
}

type TrainingCorrection struct {
	SexPosition      string        `json:"sexPosition"`
	PeoplePresent    []string      `json:"peoplePresent"`
	BodyPartsPresent []string      `json:"bodyPartsPresent"`
	ObjectsPresent   []string      `json:"objectsPresent"`
	ClothingPresent  []string      `json:"clothingPresent"`
	Boxes            []TrainingBox `json:"boxes"`
}

type TrainingSample struct {
	SampleID         string             `json:"sampleId"`
	FrameURL         string             `json:"frameUrl"`
	SourceFile       string             `json:"sourceFile"`
	SourcePath       string             `json:"sourcePath,omitempty"`
	SourceSizeBytes  int64              `json:"sourceSizeBytes,omitempty"`
	Second           float64            `json:"second"`
	CreatedAt        string             `json:"createdAt"`
	UncertaintyScore float64            `json:"uncertaintyScore,omitempty"`
	Prediction       TrainingPrediction `json:"prediction"`
}

type trainingUncertainQueueItem struct {
	SampleID         string  `json:"sampleId"`
	UncertaintyScore float64 `json:"uncertaintyScore"`
	SourceFile       string  `json:"sourceFile,omitempty"`
	CreatedAt        string  `json:"createdAt,omitempty"`
}

type trainingUncertainCandidate struct {
	sample *TrainingSample
	score  float64
}

type TrainingFeedbackRequest struct {
	SampleID   string              `json:"sampleId"`
	Accepted   bool                `json:"accepted"`
	Negative   bool                `json:"negative,omitempty"`
	Correction *TrainingCorrection `json:"correction,omitempty"`
	Notes      string              `json:"notes,omitempty"`
}

type TrainingFeedbackUpdateRequest struct {
	SampleID   string              `json:"sampleId"`
	AnsweredAt string              `json:"answeredAt"`
	Accepted   bool                `json:"accepted"`
	Negative   bool                `json:"negative,omitempty"`
	Correction *TrainingCorrection `json:"correction,omitempty"`
	Notes      string              `json:"notes,omitempty"`
}

type TrainingSkipRequest struct {
	SampleID string `json:"sampleId"`
}

type TrainingAnnotation struct {
	SampleID        string              `json:"sampleId"`
	FrameURL        string              `json:"frameUrl"`
	SourceFile      string              `json:"sourceFile"`
	SourcePath      string              `json:"sourcePath,omitempty"`
	SourceSizeBytes int64               `json:"sourceSizeBytes,omitempty"`
	Second          float64             `json:"second"`
	CreatedAt       string              `json:"createdAt"`
	AnsweredAt      string              `json:"answeredAt"`
	Prediction      TrainingPrediction  `json:"prediction"`
	Accepted        bool                `json:"accepted"`
	Negative        bool                `json:"negative,omitempty"`
	Correction      *TrainingCorrection `json:"correction,omitempty"`
	Notes           string              `json:"notes,omitempty"`
}

type TrainingDetectorPrediction struct {
	Available bool          `json:"available"`
	Source    string        `json:"source,omitempty"`
	Boxes     []TrainingBox `json:"boxes"`
}

type TrainingKeypoint struct {
	Name string  `json:"name"`
	X    float64 `json:"x"`
	Y    float64 `json:"y"`
	Conf float64 `json:"conf"`
}

type TrainingPosePerson struct {
	Label            string             `json:"label,omitempty"`
	Score            float64            `json:"score"`
	Box              TrainingBox        `json:"box"`
	Keypoints        []TrainingKeypoint `json:"keypoints"`
	Quality          float64            `json:"quality,omitempty"`
	VisibleKeypoints int                `json:"visibleKeypoints,omitempty"`
	Reliable         bool               `json:"reliable,omitempty"`
}

type TrainingPosePrediction struct {
	Available   bool                 `json:"available"`
	Source      string               `json:"source,omitempty"`
	PersonCount int                  `json:"personCount"`
	Persons     []TrainingPosePerson `json:"persons"`
}

type TrainingJobStatus struct {
	Running    bool   `json:"running"`
	Progress   int    `json:"progress"`
	Step       string `json:"step"`
	Message    string `json:"message,omitempty"`
	Error      string `json:"error,omitempty"`
	StartedAt  string `json:"startedAt,omitempty"`
	FinishedAt string `json:"finishedAt,omitempty"`
	DurationMs int64  `json:"durationMs,omitempty"`
	PreviewURL string `json:"previewUrl,omitempty"`

	Stage  string `json:"stage,omitempty"`
	Epoch  int    `json:"epoch,omitempty"`
	Epochs int    `json:"epochs,omitempty"`

	MAP50   float64 `json:"map50,omitempty"`
	MAP5095 float64 `json:"map5095,omitempty"`
}

type TrainingConfidence struct {
	Score float64 `json:"score"`
	Level string  `json:"level"`
	Label string  `json:"label"`
}

type TrainingLabelStat struct {
	Label      string             `json:"label"`
	Count      int                `json:"count"`
	Confidence TrainingConfidence `json:"confidence"`
}

type TrainingStatsLabels struct {
	People       []TrainingLabelStat `json:"people"`
	SexPositions []TrainingLabelStat `json:"sexPositions"`
	BodyParts    []TrainingLabelStat `json:"bodyParts"`
	Objects      []TrainingLabelStat `json:"objects"`
	Clothing     []TrainingLabelStat `json:"clothing"`
}

type TrainingModelInfo struct {
	TrainedAt    string  `json:"trainedAt,omitempty"`
	TrainedAtMs  int64   `json:"trainedAtMs,omitempty"`
	Epochs       int     `json:"epochs,omitempty"`
	TrainSamples int     `json:"trainSamples,omitempty"`
	ValSamples   int     `json:"valSamples,omitempty"`
	Imgsz        int     `json:"imgsz,omitempty"`
	Device       string  `json:"device,omitempty"`
	MAP50        float64 `json:"map50,omitempty"`
	MAP5095      float64 `json:"map5095,omitempty"`
}

type TrainingStatsResponse struct {
	OK                     bool                `json:"ok"`
	FeedbackCount          int                 `json:"feedbackCount"`
	AcceptedCount          int                 `json:"acceptedCount"`
	CorrectedCount         int                 `json:"correctedCount"`
	NegativeCount          int                 `json:"negativeCount"`
	SampleCount            int                 `json:"sampleCount"`
	BoxCount               int                 `json:"boxCount"`
	ModelAvailable         bool                `json:"modelAvailable"`
	ModelInfo              *TrainingModelInfo  `json:"modelInfo,omitempty"`
	DetectorModelAvailable bool                `json:"detectorModelAvailable"`
	DetectorModelInfo      *TrainingModelInfo  `json:"detectorModelInfo,omitempty"`
	PoseModelAvailable     bool                `json:"poseModelAvailable"`
	PoseModelInfo          *TrainingModelInfo  `json:"poseModelInfo,omitempty"`
	VideoMAEModelAvailable bool                `json:"videoMAEModelAvailable"`
	VideoMAEModelInfo      *TrainingModelInfo  `json:"videoMAEModelInfo,omitempty"`
	Confidence             TrainingConfidence  `json:"confidence"`
	Labels                 TrainingStatsLabels `json:"labels"`
}

type trainingProgressEvent struct {
	Type     string   `json:"type"`
	Stage    string   `json:"stage"`
	Progress float64  `json:"progress"` // 0..1
	Message  string   `json:"message,omitempty"`
	Epoch    int      `json:"epoch,omitempty"`
	Epochs   int      `json:"epochs,omitempty"`
	SampleID string   `json:"sampleId,omitempty"`
	MAP50    *float64 `json:"mAP50,omitempty"`
	MAP5095  *float64 `json:"mAP5095,omitempty"`
}

type TrainingFeedbackListResponse struct {
	OK      bool                 `json:"ok"`
	Items   []TrainingAnnotation `json:"items"`
	Total   int                  `json:"total"`
	Limit   int                  `json:"limit"`
	Offset  int                  `json:"offset"`
	HasMore bool                 `json:"hasMore"`
}

func trainingFeedbackListHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodGet {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	limit := 30
	offset := 0

	if raw := strings.TrimSpace(r.URL.Query().Get("limit")); raw != "" {
		if n, err := strconv.Atoi(raw); err == nil {
			limit = n
		}
	}

	if raw := strings.TrimSpace(r.URL.Query().Get("offset")); raw != "" {
		if n, err := strconv.Atoi(raw); err == nil {
			offset = n
		}
	}

	if limit < 1 {
		limit = 30
	}
	if limit > 200 {
		limit = 200
	}
	if offset < 0 {
		offset = 0
	}

	items, err := trainingReadAnnotations(root)
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	// Neueste zuerst.
	sort.SliceStable(items, func(i, j int) bool {
		ai := strings.TrimSpace(items[i].AnsweredAt)
		aj := strings.TrimSpace(items[j].AnsweredAt)

		if ai == aj {
			return items[i].CreatedAt > items[j].CreatedAt
		}

		return ai > aj
	})

	query := strings.TrimSpace(r.URL.Query().Get("q"))
	filter := strings.TrimSpace(r.URL.Query().Get("filter"))

	items = trainingFilterAnnotations(items, query, filter)

	total := len(items)

	if offset > total {
		offset = total
	}

	end := offset + limit
	if end > total {
		end = total
	}

	page := items[offset:end]

	trainingWriteJSON(w, http.StatusOK, TrainingFeedbackListResponse{
		OK:      true,
		Items:   page,
		Total:   total,
		Limit:   limit,
		Offset:  offset,
		HasMore: end < total,
	})
}

func trainingReadAnnotations(root string) ([]TrainingAnnotation, error) {
	path := filepath.Join(root, "feedback.jsonl")

	b, err := os.ReadFile(path)
	if err != nil {
		if os.IsNotExist(err) {
			return []TrainingAnnotation{}, nil
		}

		return nil, err
	}

	items := []TrainingAnnotation{}

	for _, line := range strings.Split(string(b), "\n") {
		line = strings.TrimSpace(line)
		if line == "" {
			continue
		}

		var item TrainingAnnotation
		if err := json.Unmarshal([]byte(line), &item); err != nil {
			continue
		}

		// Alte Einträge robust machen.
		if strings.TrimSpace(item.FrameURL) == "" && strings.TrimSpace(item.SampleID) != "" {
			item.FrameURL = "/api/training/frame?id=" + item.SampleID
		}

		items = append(items, item)
	}

	return items, nil
}

func trainingFilterAnnotations(
	items []TrainingAnnotation,
	query string,
	filter string,
) []TrainingAnnotation {
	cleanQuery := strings.ToLower(strings.TrimSpace(query))
	cleanFilter := strings.ToLower(strings.TrimSpace(filter))

	out := make([]TrainingAnnotation, 0, len(items))

	for _, item := range items {
		switch cleanFilter {
		case "accepted":
			if !item.Accepted || item.Negative {
				continue
			}

		case "corrected":
			if item.Accepted || item.Negative {
				continue
			}

		case "negative":
			if !item.Negative {
				continue
			}
		}

		if cleanQuery != "" && !trainingAnnotationMatchesQuery(item, cleanQuery) {
			continue
		}

		out = append(out, item)
	}

	return out
}

func trainingAnnotationMatchesQuery(item TrainingAnnotation, cleanQuery string) bool {
	effective := trainingEffectiveCorrection(item)

	parts := []string{
		item.SampleID,
		item.SourceFile,
		item.SourcePath,
		item.CreatedAt,
		item.AnsweredAt,
		item.Notes,
		effective.SexPosition,
	}
	if item.Negative {
		parts = append(parts, "negative negativ leer keine labels")
	}

	parts = append(parts, effective.PeoplePresent...)
	parts = append(parts, effective.BodyPartsPresent...)
	parts = append(parts, effective.ObjectsPresent...)
	parts = append(parts, effective.ClothingPresent...)

	for _, box := range effective.Boxes {
		parts = append(parts, box.Label)
	}

	haystack := strings.ToLower(strings.Join(parts, " "))

	return strings.Contains(haystack, cleanQuery)
}

func trainingRemoveSampleFromUncertainQueue(root string, sampleID string) error {
	sampleID = strings.TrimSpace(sampleID)
	if sampleID == "" {
		return nil
	}

	items, err := trainingReadUncertainQueue(root)
	if err != nil {
		return err
	}

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

	next := make([]trainingUncertainQueueItem, 0, len(items))
	changed := false

	for _, item := range items {
		if strings.TrimSpace(item.SampleID) == sampleID {
			changed = true
			continue
		}

		next = append(next, item)
	}

	if !changed {
		return nil
	}

	return trainingWriteUncertainQueue(root, next)
}

func trainingReadValidUncertainCandidates(root string) ([]trainingUncertainCandidate, error) {
	items, err := trainingReadUncertainQueue(root)
	if err != nil {
		return nil, err
	}

	if len(items) == 0 {
		return []trainingUncertainCandidate{}, nil
	}

	answered, err := trainingAnsweredSampleIDs(root)
	if err != nil {
		return nil, err
	}

	candidates := make([]trainingUncertainCandidate, 0, len(items))

	for _, item := range items {
		id := strings.TrimSpace(item.SampleID)
		if id == "" || answered[id] {
			continue
		}

		framePath := filepath.Join(root, "frames", id+".jpg")
		if !fileExistsNonEmpty(framePath) {
			continue
		}

		sample, err := trainingReadSample(root, id)
		if err != nil {
			continue
		}

		score := item.UncertaintyScore
		if score <= 0 {
			score = sample.UncertaintyScore
		}
		if score <= 0 {
			score = trainingPredictionUncertaintyScore(sample.Prediction)
		}

		score = clamp01(score)
		sample.UncertaintyScore = score

		candidates = append(candidates, trainingUncertainCandidate{
			sample: sample,
			score:  score,
		})
	}

	sort.Slice(candidates, func(i, j int) bool {
		if candidates[i].score == candidates[j].score {
			return candidates[i].sample.CreatedAt < candidates[j].sample.CreatedAt
		}

		return candidates[i].score > candidates[j].score
	})

	return candidates, nil
}

func trainingWriteUncertainCandidateQueue(root string, candidates []trainingUncertainCandidate) error {
	items := make([]trainingUncertainQueueItem, 0, len(candidates))

	for _, candidate := range candidates {
		if candidate.sample == nil {
			continue
		}

		id := strings.TrimSpace(candidate.sample.SampleID)
		if id == "" {
			continue
		}

		items = append(items, trainingUncertainQueueItem{
			SampleID:         id,
			UncertaintyScore: clamp01(candidate.score),
			SourceFile:       candidate.sample.SourceFile,
			CreatedAt:        candidate.sample.CreatedAt,
		})
	}

	return trainingWriteUncertainQueue(root, items)
}

func trainingCreateUncertainCandidateWithProgress(
	root string,
	startedAtMs int64,
	requestID string,
	stepStart int,
	stepTotal int,
	prefix string,
) (*trainingUncertainCandidate, error) {
	sample, err := trainingCreateNextSampleWithProgressRange(
		startedAtMs,
		requestID,
		stepStart,
		stepTotal,
		prefix,
	)
	if err != nil {
		return nil, err
	}

	score := trainingPredictionUncertaintyScore(sample.Prediction)
	score = clamp01(score)

	sample.UncertaintyScore = score

	if err := trainingWriteSample(root, sample); err != nil {
		return nil, err
	}

	return &trainingUncertainCandidate{
		sample: sample,
		score:  score,
	}, nil
}

func trainingUncertainQueuePath(root string) string {
	return filepath.Join(root, "uncertain_queue.json")
}

func trainingReadUncertainQueue(root string) ([]trainingUncertainQueueItem, error) {
	path := trainingUncertainQueuePath(root)

	b, err := os.ReadFile(path)
	if err != nil {
		if os.IsNotExist(err) {
			return []trainingUncertainQueueItem{}, nil
		}

		return nil, err
	}

	var items []trainingUncertainQueueItem
	if err := json.Unmarshal(b, &items); err != nil {
		return []trainingUncertainQueueItem{}, nil
	}

	return items, nil
}

func trainingWriteUncertainQueue(root string, items []trainingUncertainQueueItem) error {
	path := trainingUncertainQueuePath(root)

	if len(items) == 0 {
		_ = os.Remove(path)
		return nil
	}

	b, err := json.MarshalIndent(items, "", "  ")
	if err != nil {
		return err
	}

	return os.WriteFile(path, b, 0644)
}

func trainingPopQueuedUncertainSample(root string) (*TrainingSample, bool, error) {
	items, err := trainingReadUncertainQueue(root)
	if err != nil {
		return nil, false, err
	}

	if len(items) == 0 {
		return nil, false, nil
	}

	answered, err := trainingAnsweredSampleIDs(root)
	if err != nil {
		return nil, false, err
	}

	remaining := make([]trainingUncertainQueueItem, 0, len(items))

	for index, item := range items {
		id := strings.TrimSpace(item.SampleID)
		if id == "" || answered[id] {
			continue
		}

		framePath := filepath.Join(root, "frames", id+".jpg")
		if !fileExistsNonEmpty(framePath) {
			continue
		}

		sample, err := trainingReadSample(root, id)
		if err != nil {
			continue
		}

		sample.UncertaintyScore = item.UncertaintyScore

		remaining = append(remaining, items[index+1:]...)

		if err := trainingWriteUncertainQueue(root, remaining); err != nil {
			return nil, false, err
		}

		return sample, true, nil
	}

	_ = trainingWriteUncertainQueue(root, []trainingUncertainQueueItem{})
	return nil, false, nil
}

func trainingScaleProgress(local float64, start int, end int) int {
	if math.IsNaN(local) || math.IsInf(local, 0) {
		local = 0
	}

	local = clamp01(local)

	if end < start {
		end = start
	}

	return start + int(math.Round(local*float64(end-start)))
}

func trainingHandleProgressLine(line string, start int, end int, defaultStep string) bool {
	line = strings.TrimSpace(line)
	if line == "" {
		return false
	}

	var ev trainingProgressEvent
	if err := json.Unmarshal([]byte(line), &ev); err != nil {
		return false
	}

	if ev.Type != "progress" {
		return false
	}

	progress := trainingScaleProgress(ev.Progress, start, end)
	step := strings.TrimSpace(ev.Message)
	if step == "" {
		step = defaultStep
	}

	trainingSetJobStatus(func(s *TrainingJobStatus) {
		if progress > s.Progress {
			s.Progress = progress
		}

		s.Step = step

		if strings.TrimSpace(ev.Stage) != "" {
			s.Stage = strings.TrimSpace(ev.Stage)
		}

		if ev.Epoch > 0 {
			s.Epoch = ev.Epoch
		}

		if ev.Epochs > 0 {
			s.Epochs = ev.Epochs
		}

		if ev.MAP50 != nil && *ev.MAP50 > 0 {
			s.MAP50 = *ev.MAP50
		}
		if ev.MAP5095 != nil && *ev.MAP5095 > 0 {
			s.MAP5095 = *ev.MAP5095
		}

		sampleID := strings.TrimSpace(ev.SampleID)
		if sampleID != "" &&
			!strings.Contains(sampleID, "/") &&
			!strings.Contains(sampleID, "\\") {
			s.PreviewURL = "/api/training/frame?id=" + url.QueryEscape(sampleID)
		}
	})

	return true
}

func trainingPublishJobStatus(status TrainingJobStatus) {
	b, err := json.Marshal(map[string]any{
		"type":     "training_status",
		"training": status,
		"ts":       time.Now().UnixMilli(),
	})
	if err != nil {
		return
	}

	publishSSE("training", b)
}

func trainingPublishAnalysisStep(
	requestID string,
	startedAtMs int64,
	current int,
	total int,
	sourceFile string,
	message string,
) {
	trainingPublishAnalysisStepWithPreview(
		requestID,
		startedAtMs,
		current,
		total,
		sourceFile,
		"",
		message,
	)
}

func trainingPublishAnalysisStepWithPreview(
	requestID string,
	startedAtMs int64,
	current int,
	total int,
	sourceFile string,
	previewURL string,
	message string,
) {
	progress := 0.0
	if total > 0 {
		progress = float64(current) / float64(total)
	}

	payload := map[string]any{
		"type":        "analysis_progress",
		"scope":       "training",
		"requestId":   requestID,
		"running":     true,
		"phase":       "running",
		"progress":    progress,
		"startedAtMs": startedAtMs,
		"current":     current,
		"total":       total,
		"sourceFile":  strings.TrimSpace(sourceFile),
		"message":     strings.TrimSpace(message),
		"ts":          time.Now().UnixMilli(),
	}

	if strings.TrimSpace(previewURL) != "" {
		payload["previewUrl"] = strings.TrimSpace(previewURL)
	}

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

	publishSSE("analysisProgress", b)
}

func trainingPublishAnalysisStarted(
	requestID string,
	total int,
	sourceFile string,
	message string,
) int64 {
	return trainingPublishAnalysisStartedWithPreview(
		requestID,
		total,
		sourceFile,
		"",
		message,
	)
}

func trainingPublishAnalysisStartedWithPreview(
	requestID string,
	total int,
	sourceFile string,
	previewURL string,
	message string,
) int64 {
	startedAtMs := time.Now().UnixMilli()

	payload := map[string]any{
		"type":        "analysis_progress",
		"scope":       "training",
		"requestId":   requestID,
		"running":     true,
		"phase":       "starting",
		"progress":    0,
		"startedAtMs": startedAtMs,
		"current":     0,
		"total":       total,
		"sourceFile":  strings.TrimSpace(sourceFile),
		"message":     strings.TrimSpace(message),
		"ts":          time.Now().UnixMilli(),
	}

	if strings.TrimSpace(previewURL) != "" {
		payload["previewUrl"] = strings.TrimSpace(previewURL)
	}

	b, err := json.Marshal(payload)
	if err == nil {
		publishSSE("analysisProgress", b)
	}

	return startedAtMs
}

func trainingPublishAnalysisFinished(
	requestID string,
	startedAtMs int64,
	total int,
	sourceFile string,
	message string,
) {
	finishedAtMs := time.Now().UnixMilli()
	durationMs := finishedAtMs - startedAtMs
	if durationMs < 0 {
		durationMs = 0
	}

	b, err := json.Marshal(map[string]any{
		"type":         "analysis_progress",
		"scope":        "training",
		"requestId":    requestID,
		"running":      false,
		"phase":        "done",
		"progress":     1,
		"startedAtMs":  startedAtMs,
		"finishedAtMs": finishedAtMs,
		"durationMs":   durationMs,
		"current":      total,
		"total":        total,
		"sourceFile":   strings.TrimSpace(sourceFile),
		"message":      strings.TrimSpace(message),
		"ts":           time.Now().UnixMilli(),
	})
	if err != nil {
		return
	}

	publishSSE("analysisProgress", b)
}

func trainingPublishAnalysisError(
	requestID string,
	startedAtMs int64,
	sourceFile string,
	message string,
	err error,
) {
	finishedAtMs := time.Now().UnixMilli()
	durationMs := finishedAtMs - startedAtMs
	if durationMs < 0 {
		durationMs = 0
	}

	errText := ""
	if err != nil {
		errText = err.Error()
	}

	b, marshalErr := json.Marshal(map[string]any{
		"type":         "analysis_progress",
		"scope":        "training",
		"requestId":    requestID,
		"running":      false,
		"phase":        "error",
		"progress":     0,
		"startedAtMs":  startedAtMs,
		"finishedAtMs": finishedAtMs,
		"durationMs":   durationMs,
		"sourceFile":   strings.TrimSpace(sourceFile),
		"message":      strings.TrimSpace(message),
		"error":        errText,
		"ts":           time.Now().UnixMilli(),
	})
	if marshalErr != nil {
		return
	}

	publishSSE("analysisProgress", b)
}

func trainingRunCommandStreaming(
	ctx context.Context,
	python string,
	script string,
	onLine func(line string) bool,
	args ...string,
) (string, error) {
	cmdArgs := append([]string{script}, args...)
	cmd := exec.CommandContext(ctx, python, cmdArgs...)

	cmd.SysProcAttr = &syscall.SysProcAttr{
		HideWindow:    true,
		CreationFlags: 0x08000000,
	}

	stdout, err := cmd.StdoutPipe()
	if err != nil {
		return "", err
	}

	stderr, err := cmd.StderrPipe()
	if err != nil {
		return "", err
	}

	if err := cmd.Start(); err != nil {
		return "", err
	}

	var outMu sync.Mutex
	var lines []string

	readPipe := func(scanner *bufio.Scanner) {
		scanner.Buffer(make([]byte, 0, 64*1024), 1024*1024)

		for scanner.Scan() {
			line := strings.TrimSpace(scanner.Text())
			if line == "" {
				continue
			}

			handled := false
			if onLine != nil {
				handled = onLine(line)
			}

			// Progress-Events nicht in den finalen Output übernehmen.
			if handled {
				continue
			}

			outMu.Lock()
			lines = append(lines, line)
			outMu.Unlock()
		}
	}

	var wg sync.WaitGroup
	wg.Add(2)

	go func() {
		defer wg.Done()
		readPipe(bufio.NewScanner(stdout))
	}()

	go func() {
		defer wg.Done()
		readPipe(bufio.NewScanner(stderr))
	}()

	err = cmd.Wait()
	wg.Wait()

	outMu.Lock()
	out := strings.Join(lines, "\n")
	outMu.Unlock()

	if ctx.Err() != nil {
		return strings.TrimSpace(out), errTrainingCancelled
	}

	return strings.TrimSpace(out), err
}

const minTrainingFeedbackCount = 5

const minDetectorTrainCount = 20
const minDetectorValCount = 3
const minPoseTrainCount = 20
const minPoseValCount = 3
const trainingPoseKeypointCount = 17
const trainingPoseKeypointMinConfidence = 0.20
const trainingPoseReliableMinScore = 0.30
const trainingPoseReliableMinKeypoints = 6
const trainingPoseReliableMinQuality = 0.45
const trainingPositionContextMinScore = 0.22
const trainingPositionContextMaxScore = 0.44
const trainingPositionContextBoostWeight = 0.60
const trainingPositionContextOverrideMargin = 0.16

var errTrainingCancelled = errors.New("training cancelled")

var trainingJob = struct {
	mu     sync.Mutex
	status TrainingJobStatus
	cancel context.CancelFunc
}{}

func trainingSetJobStatus(update func(*TrainingJobStatus)) {
	trainingJob.mu.Lock()
	update(&trainingJob.status)
	snapshot := trainingJob.status
	trainingJob.mu.Unlock()

	trainingPublishJobStatus(snapshot)
}

func trainingGetJobStatus() TrainingJobStatus {
	trainingJob.mu.Lock()
	defer trainingJob.mu.Unlock()
	return trainingJob.status
}

func trainingStartJob(cancel context.CancelFunc) {
	trainingJob.mu.Lock()

	trainingJob.status = TrainingJobStatus{
		Running:   true,
		Progress:  5,
		Step:      "Training wird vorbereitet…",
		StartedAt: time.Now().UTC().Format(time.RFC3339),
	}

	trainingJob.cancel = cancel
	snapshot := trainingJob.status

	trainingJob.mu.Unlock()

	trainingPublishJobStatus(snapshot)
}

func trainingClearJobCancel() {
	trainingJob.mu.Lock()
	trainingJob.cancel = nil
	trainingJob.mu.Unlock()
}

func trainingCleanupPartialDetectorTraining(root string) error {
	// Löscht nur temporäre Trainingsläufe.
	// Feedback, Samples, Frames und Dataset bleiben erhalten.
	runsDir := filepath.Join(root, "detector", "runs")

	if err := os.RemoveAll(runsDir); err != nil {
		return err
	}

	return os.MkdirAll(runsDir, 0755)
}

func trainingFinishCancelled(root string) {
	cleanupErr := trainingCleanupPartialDetectorTraining(root)

	trainingSetJobStatus(func(s *TrainingJobStatus) {
		finishedAt := time.Now().UTC()

		var durationMs int64
		if startedAt, err := time.Parse(time.RFC3339, strings.TrimSpace(s.StartedAt)); err == nil {
			durationMs = finishedAt.Sub(startedAt).Milliseconds()
			if durationMs < 0 {
				durationMs = 0
			}
		}

		s.Running = false
		s.Step = "Training abgebrochen."
		s.Message = "Training wurde abgebrochen. Temporäre Trainingsausgaben wurden gelöscht."
		s.Error = ""
		s.FinishedAt = finishedAt.Format(time.RFC3339)
		s.DurationMs = durationMs
		s.PreviewURL = ""

		if cleanupErr != nil {
			s.Message = "Training wurde abgebrochen, aber temporäre Trainingsausgaben konnten nicht vollständig gelöscht werden."
			s.Error = cleanupErr.Error()
		}
	})

	trainingClearJobCancel()
}

func trainingRunCommand(python string, script string, args ...string) (string, error) {
	cmdArgs := append([]string{script}, args...)
	cmd := exec.Command(python, cmdArgs...)

	cmd.SysProcAttr = &syscall.SysProcAttr{
		HideWindow:    true,
		CreationFlags: 0x08000000,
	}

	out, err := cmd.CombinedOutput()
	return strings.TrimSpace(string(out)), err
}

type trainingModelResolution struct {
	BestPath        string
	EffectivePath   string
	Source          string
	TrainedExists   bool
	EffectiveExists bool
}

func trainingResolveModel(
	root string,
	kind string,
	trainedSource string,
) trainingModelResolution {
	bestPath := filepath.Join(root, kind, "model", "best.pt")
	if fileExistsNonEmpty(bestPath) {
		return trainingModelResolution{
			BestPath:        bestPath,
			EffectivePath:   bestPath,
			Source:          trainedSource,
			TrainedExists:   true,
			EffectiveExists: true,
		}
	}

	return trainingModelResolution{
		BestPath:        bestPath,
		EffectivePath:   bestPath,
		Source:          kind + "_missing",
		TrainedExists:   false,
		EffectiveExists: false,
	}
}

func trainingResolveDetectorModel(root string) trainingModelResolution {
	return trainingResolveModel(
		root,
		"detector",
		"yolo26_detector",
	)
}

func trainingResolvePoseModel(root string) trainingModelResolution {
	res := trainingResolveModel(
		root,
		"pose",
		"yolo_pose",
	)
	if res.EffectiveExists {
		return res
	}

	if poseBase, err := embeddedPoseModelPath(); err == nil && fileExistsNonEmpty(poseBase) {
		res.EffectivePath = poseBase
		res.Source = "yolo26_pose_base"
		res.EffectiveExists = true
	}

	return res
}

func trainingLabelsHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodGet {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	trainingWriteJSON(w, http.StatusOK, defaultTrainingLabelsFromJSON())
}

const trainingImportVideoDefaultFrameCount = 10
const trainingImportVideoMaxFrameCount = 30

type TrainingImportVideoRequest struct {
	JobID             string `json:"jobId"`
	Output            string `json:"output"`
	Count             int    `json:"count"`
	AnalysisRequestID string `json:"analysisRequestId"`
}

type TrainingImportVideoResponse struct {
	OK      bool             `json:"ok"`
	Count   int              `json:"count"`
	Sample  *TrainingSample  `json:"sample,omitempty"`
	Samples []TrainingSample `json:"samples,omitempty"`
	Errors  []string         `json:"errors,omitempty"`
}

func trainingCleanImportVideoCount(count int) int {
	if count <= 0 {
		return trainingImportVideoDefaultFrameCount
	}

	if count > trainingImportVideoMaxFrameCount {
		return trainingImportVideoMaxFrameCount
	}

	return count
}

func trainingSupportedImportVideo(path string) bool {
	switch strings.ToLower(filepath.Ext(path)) {
	case ".mp4", ".m4v", ".mov", ".mkv", ".webm":
		return true
	default:
		return false
	}
}

func trainingGeneratedAssetIDCandidatesForVideo(videoPath string) []string {
	videoPath = strings.TrimSpace(videoPath)
	if videoPath == "" {
		return nil
	}

	out := []string{}
	seen := map[string]bool{}

	add := func(id string) {
		id = stripHotPrefix(strings.TrimSpace(id))

		if id == "" ||
			id == "." ||
			id == ".." ||
			strings.Contains(id, "/") ||
			strings.Contains(id, "\\") {
			return
		}

		if seen[id] {
			return
		}

		seen[id] = true
		out = append(out, id)
	}

	// Fall 1:
	// Video liegt selbst unter /generated/<id>/...
	//
	// Beispiel:
	// C:\app\generated\abc123\video.mp4
	// => abc123
	slashPath := filepath.ToSlash(filepath.Clean(videoPath))
	parts := strings.Split(slashPath, "/")

	for i := 0; i+1 < len(parts); i++ {
		if strings.EqualFold(parts[i], "generated") {
			add(parts[i+1])
		}
	}

	// Fall 2:
	// Video liegt z.B. in done/keep, aber generated/<id>/preview.jpg
	// basiert auf dem Dateinamen ohne Extension.
	//
	// Beispiel:
	// done/keep/model/abc123.mp4
	// => generated/abc123/preview.jpg
	base := filepath.Base(videoPath)
	stem := strings.TrimSuffix(base, filepath.Ext(base))
	add(stem)

	return out
}

func trainingGeneratedPreviewPathForAssetID(assetID string) (string, bool) {
	assetID = stripHotPrefix(strings.TrimSpace(assetID))

	if assetID == "" ||
		assetID == "." ||
		assetID == ".." ||
		strings.Contains(assetID, "/") ||
		strings.Contains(assetID, "\\") {
		return "", false
	}

	previewPath, err := resolvePathRelativeToApp(
		filepath.Join("generated", assetID, "preview.jpg"),
	)
	if err != nil {
		return "", false
	}

	if !fileExistsNonEmpty(previewPath) {
		return "", false
	}

	return previewPath, true
}

func trainingPreviewPathForVideo(videoPath string) (string, bool) {
	for _, assetID := range trainingGeneratedAssetIDCandidatesForVideo(videoPath) {
		if previewPath, ok := trainingGeneratedPreviewPathForAssetID(assetID); ok {
			return previewPath, true
		}
	}

	return "", false
}

func trainingPreviewURLForVideoPath(videoPath string) string {
	videoPath = strings.TrimSpace(videoPath)
	if videoPath == "" {
		return ""
	}

	if !trainingSupportedImportVideo(videoPath) {
		return ""
	}

	return "/api/training/video-preview?output=" + url.QueryEscape(videoPath)
}

func trainingPreviewAssetIDForVideo(videoPath string) string {
	candidates := trainingGeneratedAssetIDCandidatesForVideo(videoPath)

	for _, assetID := range candidates {
		if _, ok := trainingGeneratedPreviewPathForAssetID(assetID); ok {
			return assetID
		}
	}

	for _, assetID := range candidates {
		if _, err := findFinishedFileByID(assetID); err == nil {
			return assetID
		}
	}

	if len(candidates) > 0 {
		return candidates[0]
	}

	return ""
}

func trainingVideoPreviewHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodGet && r.Method != http.MethodHead {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	outPath := strings.TrimSpace(r.URL.Query().Get("output"))
	if outPath == "" {
		trainingWriteError(w, http.StatusBadRequest, "output missing")
		return
	}

	if !trainingSupportedImportVideo(outPath) {
		trainingWriteError(w, http.StatusBadRequest, "unsupported video type")
		return
	}

	st, err := os.Stat(outPath)
	if err != nil || st == nil || st.IsDir() || st.Size() <= 0 {
		trainingWriteError(w, http.StatusNotFound, "video not found")
		return
	}

	// Fast path: Wenn /generated/<id>/preview.jpg schon existiert, direkt ausliefern.
	if previewPath, ok := trainingPreviewPathForVideo(outPath); ok {
		w.Header().Set("Cache-Control", "no-store")
		servePreviewJPGFile(w, r, previewPath)
		return
	}

	assetID := trainingPreviewAssetIDForVideo(outPath)
	if assetID == "" {
		trainingWriteError(w, http.StatusNotFound, "preview asset id not found")
		return
	}

	// Wichtig:
	// Nicht file=preview.jpg setzen.
	// Ohne file=... darf recordPreviewWithBase die Preview bei Bedarf erzeugen.
	r2 := r.Clone(r.Context())
	u := *r.URL
	q := u.Query()

	q.Set("id", assetID)
	q.Del("output")
	q.Del("file")
	q.Del("fallbackOnly")

	u.RawQuery = q.Encode()
	r2.URL = &u

	recordPreviewWithBase(w, r2, "/api/training/video-preview")
}

func trainingFrameSecondsForVideo(duration float64, count int) []float64 {
	count = trainingCleanImportVideoCount(count)

	if duration <= 0 {
		return []float64{0}
	}

	if duration <= 2 {
		return []float64{0}
	}

	minSec := 0.5
	maxSec := duration - 0.5

	if maxSec <= minSec {
		return []float64{0}
	}

	out := make([]float64, 0, count)

	// Lokaler RNG, damit jeder Import neue Frames bekommt.
	rng := rand.New(rand.NewSource(time.Now().UnixNano()))

	// Mindestabstand zwischen Frames, damit nicht mehrfach fast dieselbe Stelle kommt.
	minDistance := 0.4

	// Bei sehr kurzen Videos Abstand automatisch verkleinern.
	availableRange := maxSec - minSec
	if availableRange/float64(count) < minDistance {
		minDistance = math.Max(0.1, availableRange/float64(count+1))
	}

	const maxAttempts = 4000

	for attempts := 0; len(out) < count && attempts < maxAttempts; attempts++ {
		sec := minSec + rng.Float64()*(maxSec-minSec)

		// Auf 0.1s runden, damit IDs/Logs lesbar bleiben.
		sec = math.Round(sec*10) / 10

		tooClose := false
		for _, existing := range out {
			if math.Abs(sec-existing) < minDistance {
				tooClose = true
				break
			}
		}

		if tooClose {
			continue
		}

		out = append(out, sec)
	}

	// Fallback, falls ein extrem kurzes Video nicht genug unterschiedliche Random-Punkte hergibt.
	for len(out) < count {
		sec := minSec + rng.Float64()*(maxSec-minSec)
		sec = math.Round(sec*10) / 10
		out = append(out, sec)
	}

	sort.Float64s(out)

	if len(out) == 0 {
		out = append(out, 0)
	}

	return out
}

func trainingImportVideoHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodPost {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	var req TrainingImportVideoRequest
	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
		trainingWriteError(w, http.StatusBadRequest, "invalid json")
		return
	}

	outPath := strings.TrimSpace(req.Output)
	if outPath == "" {
		trainingWriteError(w, http.StatusBadRequest, "output missing")
		return
	}

	if !trainingSupportedImportVideo(outPath) {
		trainingWriteError(w, http.StatusBadRequest, "unsupported video type")
		return
	}

	fi, err := os.Stat(outPath)
	if err != nil || fi == nil || fi.IsDir() || fi.Size() <= 0 {
		if err == nil {
			err = errors.New("video file missing or empty")
		}

		trainingWriteError(w, http.StatusBadRequest, "video not found: "+err.Error())
		return
	}

	duration := trainingProbeDurationSeconds(outPath)
	if duration <= 0 {
		trainingWriteError(w, http.StatusBadRequest, "Videolänge konnte nicht bestimmt werden")
		return
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	if err := trainingEnsureDetectorDirs(root); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	if err := os.MkdirAll(filepath.Join(root, "frames"), 0755); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	if err := os.MkdirAll(filepath.Join(root, "samples"), 0755); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	seconds := trainingFrameSecondsForVideo(duration, req.Count)
	sourceFile := filepath.Base(outPath)
	previewURL := ""

	requestID := strings.TrimSpace(req.AnalysisRequestID)
	if requestID == "" {
		requestID = trainingMakeSampleID(outPath, float64(time.Now().UnixNano()))
	}

	totalSteps := len(seconds) * 3
	if totalSteps < 1 {
		totalSteps = 1
	}

	startedAtMs := trainingPublishAnalysisStartedWithPreview(
		requestID,
		totalSteps,
		sourceFile,
		previewURL,
		"Video wird ins Training übernommen…",
	)

	var sourceSizeBytes int64
	if st, err := os.Stat(outPath); err == nil && st != nil && !st.IsDir() {
		sourceSizeBytes = st.Size()
	}

	samples := make([]TrainingSample, 0, len(seconds))
	errs := []string{}

	for i, second := range seconds {
		stepBase := i * 3

		trainingPublishAnalysisStepWithPreview(
			requestID,
			startedAtMs,
			stepBase+1,
			totalSteps,
			sourceFile,
			previewURL,
			fmt.Sprintf("Frame %d/%d wird extrahiert…", i+1, len(seconds)),
		)

		id := trainingMakeSampleID(outPath, second)
		framePath := filepath.Join(root, "frames", id+".jpg")

		if err := trainingExtractFrame(outPath, framePath, second); err != nil {
			errs = append(errs, fmt.Sprintf("Frame bei %.1fs: %v", second, err))
			continue
		}

		// Nach jeder erfolgreichen Extraktion das aktuell verarbeitete Frame zeigen.
		previewURL = "/api/training/frame?id=" + url.QueryEscape(id)

		trainingPublishAnalysisStepWithPreview(
			requestID,
			startedAtMs,
			stepBase+2,
			totalSteps,
			sourceFile,
			previewURL,
			fmt.Sprintf("Frame %d/%d wird analysiert…", i+1, len(seconds)),
		)

		prediction := trainingPredictFrame(framePath)

		sample := &TrainingSample{
			SampleID:        id,
			FrameURL:        "/api/training/frame?id=" + id,
			SourceFile:      sourceFile,
			SourcePath:      outPath,
			SourceSizeBytes: sourceSizeBytes,
			Second:          second,
			CreatedAt:       time.Now().UTC().Format(time.RFC3339),
			Prediction:      prediction,
		}

		trainingPublishAnalysisStepWithPreview(
			requestID,
			startedAtMs,
			stepBase+3,
			totalSteps,
			sourceFile,
			previewURL,
			fmt.Sprintf("Frame %d/%d wird gespeichert…", i+1, len(seconds)),
		)

		if err := trainingWriteSample(root, sample); err != nil {
			_ = os.Remove(framePath)
			errs = append(errs, fmt.Sprintf("Frame bei %.1fs speichern: %v", second, err))
			continue
		}

		samples = append(samples, *sample)
	}

	if len(samples) == 0 {
		msg := "keine Trainingsframes erzeugt"
		if len(errs) > 0 {
			msg += ": " + strings.Join(errs, "; ")
		}

		err := errors.New(msg)

		trainingPublishAnalysisError(
			requestID,
			startedAtMs,
			sourceFile,
			"Video konnte nicht ins Training übernommen werden.",
			err,
		)

		trainingWriteError(w, http.StatusInternalServerError, msg)
		return
	}

	trainingPublishAnalysisFinished(
		requestID,
		startedAtMs,
		totalSteps,
		sourceFile,
		fmt.Sprintf("%d Frames ins Training übernommen.", len(samples)),
	)

	trainingWriteJSON(w, http.StatusOK, TrainingImportVideoResponse{
		OK:      true,
		Count:   len(samples),
		Sample:  &samples[0],
		Samples: samples,
		Errors:  errs,
	})
}

func trainingNextHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodGet {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	forceNew := r.URL.Query().Get("force") == "1" ||
		strings.EqualFold(r.URL.Query().Get("force"), "true")

	analysisRequestID := strings.TrimSpace(r.URL.Query().Get("analysisRequestId"))

	excludeIDs := trainingExcludedSampleIDs(r)

	preferUncertain := strings.EqualFold(r.URL.Query().Get("mode"), "uncertain") ||
		strings.EqualFold(r.URL.Query().Get("sampleMode"), "uncertain")

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	refreshPrediction := r.URL.Query().Get("refresh") == "1" ||
		strings.EqualFold(r.URL.Query().Get("refresh"), "true")

	if !forceNew && !preferUncertain {
		var startedAtMs int64

		if refreshPrediction {
			startedAtMs = time.Now().UnixMilli()
			trainingPublishAnalysisStep(
				analysisRequestID,
				startedAtMs,
				1,
				2,
				"",
				"Aktuelles Bild wird neu analysiert…",
			)
		}

		if sample, ok, err := trainingLatestOpenSample(root, refreshPrediction, startedAtMs, analysisRequestID, excludeIDs); err != nil {
			if refreshPrediction {
				trainingPublishAnalysisError(
					analysisRequestID,
					startedAtMs,
					"",
					"Aktuelles Bild konnte nicht neu analysiert werden.",
					err,
				)
			}

			trainingWriteError(w, http.StatusInternalServerError, err.Error())
			return
		} else if ok {
			if refreshPrediction {
				trainingPublishAnalysisFinished(
					analysisRequestID,
					startedAtMs,
					2,
					sample.SourceFile,
					"Analyse abgeschlossen.",
				)
			}

			trainingWriteJSON(w, http.StatusOK, sample)
			return
		}
	}

	startedAtMs := trainingPublishAnalysisStarted(
		analysisRequestID,
		func() int {
			if preferUncertain {
				return trainingUncertainCandidateCount*4 + 1
			}

			return 4
		}(),
		"",
		func() string {
			if preferUncertain {
				return "Unsichere Prediction wird gesucht…"
			}

			return "Neues Trainingsbild wird vorbereitet…"
		}(),
	)

	var sample *TrainingSample

	if preferUncertain {
		sample, err = trainingCreateUncertainNextSampleWithProgress(startedAtMs, analysisRequestID)
	} else {
		sample, err = trainingCreateNextSampleWithProgress(startedAtMs, analysisRequestID)
	}

	if err != nil {
		trainingPublishAnalysisError(
			analysisRequestID,
			startedAtMs,
			"",
			"Trainingsbild konnte nicht erstellt werden.",
			err,
		)
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	trainingPublishAnalysisFinished(
		analysisRequestID,
		startedAtMs,
		4,
		sample.SourceFile,
		"Analyse abgeschlossen.",
	)

	trainingWriteJSON(w, http.StatusOK, sample)
}

func trainingLatestOpenSample(
	root string,
	refreshPrediction bool,
	startedAtMs int64,
	requestID string,
	excludeIDs map[string]bool,
) (*TrainingSample, bool, error) {
	answered, err := trainingAnsweredSampleIDs(root)
	if err != nil {
		return nil, false, err
	}

	samplesDir := filepath.Join(root, "samples")

	entries, err := os.ReadDir(samplesDir)
	if err != nil {
		if os.IsNotExist(err) {
			return nil, false, nil
		}
		return nil, false, err
	}

	type sampleFile struct {
		id      string
		path    string
		modTime time.Time
	}

	files := []sampleFile{}

	for _, entry := range entries {
		if entry.IsDir() {
			continue
		}

		name := entry.Name()
		if strings.ToLower(filepath.Ext(name)) != ".json" {
			continue
		}

		id := strings.TrimSuffix(name, filepath.Ext(name))
		if id == "" || answered[id] || excludeIDs[id] {
			continue
		}

		info, err := entry.Info()
		if err != nil {
			continue
		}

		files = append(files, sampleFile{
			id:      id,
			path:    filepath.Join(samplesDir, name),
			modTime: info.ModTime(),
		})
	}

	sort.Slice(files, func(i, j int) bool {
		return files[i].modTime.After(files[j].modTime)
	})

	for _, file := range files {
		sample, err := trainingReadSample(root, file.id)
		if err != nil {
			continue
		}

		framePath := filepath.Join(root, "frames", sample.SampleID+".jpg")
		if !fileExistsNonEmpty(framePath) {
			continue
		}

		if refreshPrediction {
			sourceFile := strings.TrimSpace(sample.SourceFile)
			if sourceFile == "" {
				sourceFile = filepath.Base(sample.SourcePath)
			}

			trainingPublishAnalysisStep(
				requestID,
				startedAtMs,
				1,
				2,
				sourceFile,
				"Aktuelles Bild wird analysiert…",
			)

			sample.Prediction = trainingPredictFrame(framePath)

			trainingPublishAnalysisStep(
				requestID,
				startedAtMs,
				2,
				2,
				sourceFile,
				"Analyse-Ergebnis wird gespeichert…",
			)

			if err := trainingWriteSample(root, sample); err != nil {
				return nil, false, err
			}
		}

		return sample, true, nil
	}

	return nil, false, nil
}

func trainingExcludedSampleIDs(r *http.Request) map[string]bool {
	out := map[string]bool{}

	for _, raw := range r.URL.Query()["exclude"] {
		for _, part := range strings.Split(raw, ",") {
			id := strings.TrimSpace(part)
			if id == "" {
				continue
			}

			if strings.Contains(id, "/") || strings.Contains(id, "\\") {
				continue
			}

			out[id] = true
		}
	}

	return out
}

func trainingAnsweredSampleIDs(root string) (map[string]bool, error) {
	out := map[string]bool{}

	path := filepath.Join(root, "feedback.jsonl")

	b, err := os.ReadFile(path)
	if err != nil {
		if os.IsNotExist(err) {
			return out, nil
		}
		return nil, err
	}

	for _, line := range strings.Split(string(b), "\n") {
		line = strings.TrimSpace(line)
		if line == "" {
			continue
		}

		var row TrainingAnnotation
		if err := json.Unmarshal([]byte(line), &row); err != nil {
			continue
		}

		id := strings.TrimSpace(row.SampleID)
		if id != "" {
			out[id] = true
		}
	}

	return out, nil
}

func trainingFrameHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodGet {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	id := strings.TrimSpace(r.URL.Query().Get("id"))
	if id == "" || strings.Contains(id, "/") || strings.Contains(id, "\\") {
		trainingWriteError(w, http.StatusBadRequest, "invalid frame id")
		return
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	path := filepath.Join(root, "frames", id+".jpg")

	if _, err := os.Stat(path); err != nil {
		trainingWriteError(w, http.StatusNotFound, "frame not found")
		return
	}

	w.Header().Set("Cache-Control", "no-store")
	http.ServeFile(w, r, path)
}

func trainingDetectorBoxesForAnnotation(sample *TrainingSample, req TrainingFeedbackRequest) []TrainingBox {
	if req.Negative {
		return []TrainingBox{}
	}

	boxes := []TrainingBox{}

	if req.Correction != nil {
		boxes = append(boxes, req.Correction.Boxes...)
	} else if req.Accepted {
		boxes = append(boxes, sample.Prediction.Boxes...)
	}

	return boxes
}

func trainingNegativeCorrection() *TrainingCorrection {
	return &TrainingCorrection{
		SexPosition:      trainingNoSexPositionLabel,
		PeoplePresent:    []string{},
		BodyPartsPresent: []string{},
		ObjectsPresent:   []string{},
		ClothingPresent:  []string{},
		Boxes:            []TrainingBox{},
	}
}

func trainingFeedbackHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodPost {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	var req TrainingFeedbackRequest
	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
		trainingWriteError(w, http.StatusBadRequest, "invalid json")
		return
	}

	req.SampleID = strings.TrimSpace(req.SampleID)
	if req.SampleID == "" {
		trainingWriteError(w, http.StatusBadRequest, "sampleId missing")
		return
	}
	if req.Negative {
		req.Accepted = false
		req.Correction = trainingNegativeCorrection()
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	sample, err := trainingReadSample(root, req.SampleID)
	if err != nil {
		trainingWriteError(w, http.StatusNotFound, "sample not found")
		return
	}

	annotation := TrainingAnnotation{
		SampleID:        sample.SampleID,
		FrameURL:        sample.FrameURL,
		SourceFile:      sample.SourceFile,
		SourcePath:      sample.SourcePath,
		SourceSizeBytes: sample.SourceSizeBytes,
		Second:          sample.Second,
		CreatedAt:       sample.CreatedAt,
		AnsweredAt:      time.Now().UTC().Format(time.RFC3339),
		Prediction:      sample.Prediction,
		Accepted:        req.Accepted,
		Negative:        req.Negative,
		Correction:      req.Correction,
		Notes:           strings.TrimSpace(req.Notes),
	}

	if !annotation.Accepted && annotation.Correction == nil {
		trainingWriteError(w, http.StatusBadRequest, "correction missing")
		return
	}

	if err := trainingAppendAnnotation(root, annotation); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	detectorBoxes := trainingDetectorBoxesForAnnotation(sample, req)

	if req.Negative || len(detectorBoxes) > 0 {
		if err := trainingWriteDetectorSample(root, sample, detectorBoxes, req.Negative); err != nil {
			appLogln("⚠️ detector sample write failed:", err)
		}
	}

	trainingDeletePoseSample(root, req.SampleID)
	if sexPosition := trainingSexPositionForFeedback(sample, req); !isNoSexPositionLabel(sexPosition) {
		if err := trainingWritePoseSample(root, sample, sexPosition, detectorBoxes); err != nil {
			appLogln("pose sample write failed:", err)
		}
	}

	trainingWriteJSON(w, http.StatusOK, map[string]any{
		"ok": true,
	})
}

func trainingFeedbackUpdateHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodPut && r.Method != http.MethodPost {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	var req TrainingFeedbackUpdateRequest
	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
		trainingWriteError(w, http.StatusBadRequest, "invalid json")
		return
	}

	req.SampleID = strings.TrimSpace(req.SampleID)
	req.AnsweredAt = strings.TrimSpace(req.AnsweredAt)
	if req.Negative {
		req.Accepted = false
		req.Correction = trainingNegativeCorrection()
	}

	if req.SampleID == "" {
		trainingWriteError(w, http.StatusBadRequest, "sampleId missing")
		return
	}

	if req.AnsweredAt == "" {
		trainingWriteError(w, http.StatusBadRequest, "answeredAt missing")
		return
	}

	if strings.Contains(req.SampleID, "/") || strings.Contains(req.SampleID, "\\") {
		trainingWriteError(w, http.StatusBadRequest, "invalid sampleId")
		return
	}

	if !req.Accepted && req.Correction == nil {
		trainingWriteError(w, http.StatusBadRequest, "correction missing")
		return
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	items, err := trainingReadAnnotations(root)
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	matchIndex := -1

	for i, item := range items {
		if strings.TrimSpace(item.SampleID) == req.SampleID &&
			strings.TrimSpace(item.AnsweredAt) == req.AnsweredAt {
			matchIndex = i
			break
		}
	}

	if matchIndex < 0 {
		trainingWriteError(w, http.StatusNotFound, "feedback not found")
		return
	}

	old := items[matchIndex]

	updated := old
	updated.Accepted = req.Accepted
	updated.Negative = req.Negative
	updated.Notes = strings.TrimSpace(req.Notes)

	if req.Accepted {
		updated.Correction = nil
	} else {
		updated.Correction = req.Correction
	}

	items[matchIndex] = updated

	if err := trainingWriteAnnotations(root, items); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	sample, sampleErr := trainingReadSample(root, req.SampleID)
	if sampleErr != nil {
		sample = &TrainingSample{
			SampleID:        old.SampleID,
			FrameURL:        old.FrameURL,
			SourceFile:      old.SourceFile,
			SourcePath:      old.SourcePath,
			SourceSizeBytes: old.SourceSizeBytes,
			Second:          old.Second,
			CreatedAt:       old.CreatedAt,
			Prediction:      old.Prediction,
		}
	}

	trainingDeleteDetectorSample(root, req.SampleID)

	detectorBoxes := trainingDetectorBoxesForAnnotation(sample, TrainingFeedbackRequest{
		SampleID:   req.SampleID,
		Accepted:   req.Accepted,
		Negative:   req.Negative,
		Correction: req.Correction,
		Notes:      req.Notes,
	})

	if req.Negative || len(detectorBoxes) > 0 {
		if err := trainingWriteDetectorSample(root, sample, detectorBoxes, req.Negative); err != nil {
			appLogln("⚠️ detector sample update failed:", err)
		}
	}

	trainingDeletePoseSample(root, req.SampleID)
	if sexPosition := trainingSexPositionForFeedback(sample, TrainingFeedbackRequest{
		SampleID:   req.SampleID,
		Accepted:   req.Accepted,
		Negative:   req.Negative,
		Correction: req.Correction,
		Notes:      req.Notes,
	}); !isNoSexPositionLabel(sexPosition) {
		if err := trainingWritePoseSample(root, sample, sexPosition, detectorBoxes); err != nil {
			appLogln("pose sample update failed:", err)
		}
	}

	trainingWriteJSON(w, http.StatusOK, map[string]any{
		"ok":   true,
		"item": updated,
	})
}

func trainingSkipHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodPost && r.Method != http.MethodDelete {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	var req TrainingSkipRequest
	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
		trainingWriteError(w, http.StatusBadRequest, "invalid json")
		return
	}

	sampleID := strings.TrimSpace(req.SampleID)
	if sampleID == "" {
		trainingWriteError(w, http.StatusBadRequest, "sampleId missing")
		return
	}

	if strings.Contains(sampleID, "/") || strings.Contains(sampleID, "\\") {
		trainingWriteError(w, http.StatusBadRequest, "invalid sampleId")
		return
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	// Aus Uncertain-Queue entfernen, falls es dort noch liegt.
	if err := trainingRemoveSampleFromUncertainQueue(root, sampleID); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	// Sample + Frame löschen.
	trainingDeleteSampleFiles(root, sampleID)

	trainingWriteJSON(w, http.StatusOK, map[string]any{
		"ok":       true,
		"sampleId": sampleID,
	})
}

func trainingHasDetectorTrainingData(imagesDir string, labelsDir string) bool {
	imageExts := map[string]bool{
		".jpg":  true,
		".jpeg": true,
		".png":  true,
		".webp": true,
	}

	entries, err := os.ReadDir(imagesDir)
	if err != nil {
		return false
	}

	count := 0

	for _, e := range entries {
		if e.IsDir() {
			continue
		}

		ext := strings.ToLower(filepath.Ext(e.Name()))
		if !imageExts[ext] {
			continue
		}

		id := strings.TrimSuffix(e.Name(), filepath.Ext(e.Name()))
		labelPath := filepath.Join(labelsDir, id+".txt")

		if fileExists(labelPath) {
			count++
		}
	}

	return count >= 5
}

func trainingTrainHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodPost {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	current := trainingGetJobStatus()
	if current.Running {
		trainingWriteJSON(w, http.StatusOK, map[string]any{
			"ok":       true,
			"message":  "Training läuft bereits.",
			"training": current,
		})
		return
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	if err := trainingEnsureDetectorDirs(root); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	// Falls bisher alles zufällig in train gelandet ist, erzeugen wir mindestens
	// ein Validation-Sample durch Kopieren. Bei mehr Daten solltest du später
	// einen echten 80/20 Split verwenden.
	if err := trainingEnsurePoseDirs(root); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}
	if err := trainingEnsureVideoMAEDirs(root); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	if err := trainingEnsureDetectorValidationSample(root); err != nil {
		appLogln("⚠️ detector val sample ensure failed:", err)
	}

	if err := trainingEnsurePoseValidationSample(root); err != nil {
		appLogln("pose val sample ensure failed:", err)
	}

	feedbackPath := filepath.Join(root, "feedback.jsonl")
	feedbackCount, _ := trainingCountAnnotations(feedbackPath)

	if feedbackCount < minTrainingFeedbackCount {
		trainingWriteError(
			w,
			http.StatusBadRequest,
			fmt.Sprintf(
				"Zu wenige Bewertungen für das YOLO26-Training. Mindestens %d, aktuell %d.",
				minTrainingFeedbackCount,
				feedbackCount,
			),
		)
		return
	}

	detectorTrainImages := filepath.Join(root, "detector", "dataset", "images", "train")
	detectorTrainLabels := filepath.Join(root, "detector", "dataset", "labels", "train")
	detectorValImages := filepath.Join(root, "detector", "dataset", "images", "val")
	detectorValLabels := filepath.Join(root, "detector", "dataset", "labels", "val")
	detectorDatasetYAML := filepath.Join(root, "detector", "dataset", "dataset.yaml")
	poseTrainImages := filepath.Join(root, "pose", "dataset", "images", "train")
	poseTrainLabels := filepath.Join(root, "pose", "dataset", "labels", "train")
	poseValImages := filepath.Join(root, "pose", "dataset", "images", "val")
	poseValLabels := filepath.Join(root, "pose", "dataset", "labels", "val")
	poseDatasetYAML := filepath.Join(root, "pose", "dataset", "dataset.yaml")
	videoMAEManifest := trainingVideoMAEManifestPath(root)

	trainCount := trainingCountDetectorSamples(detectorTrainImages, detectorTrainLabels)
	valCount := trainingCountDetectorSamples(detectorValImages, detectorValLabels)
	positiveTrainCount := trainingCountPositiveDetectorSamples(detectorTrainImages, detectorTrainLabels)
	positiveValCount := trainingCountPositiveDetectorSamples(detectorValImages, detectorValLabels)
	poseTrainCount := trainingCountDetectorSamples(poseTrainImages, poseTrainLabels)
	poseValCount := trainingCountDetectorSamples(poseValImages, poseValLabels)
	videoMAETrainCount, videoMAEValCount := trainingCountVideoMAEManifestSamples(root)
	videoMAEEligibleCount, _ := trainingCountVideoMAEEligibleAnnotations(root)

	detectorDataReady := fileExistsNonEmpty(detectorDatasetYAML) &&
		trainCount >= minDetectorTrainCount &&
		valCount >= minDetectorValCount &&
		positiveTrainCount > 0 &&
		positiveValCount > 0
	poseDataReady := fileExistsNonEmpty(poseDatasetYAML) &&
		poseTrainCount >= minPoseTrainCount &&
		poseValCount >= minPoseValCount
	videoMAEDataReady := videoMAEEligibleCount >= minVideoMAETrainCount ||
		(videoMAETrainCount >= minVideoMAETrainCount && videoMAEValCount >= minVideoMAEValCount)

	if detectorDataReady || poseDataReady || videoMAEDataReady {
		goto startTraining
	}

	if !fileExistsNonEmpty(detectorDatasetYAML) ||
		trainCount < minDetectorTrainCount ||
		valCount < minDetectorValCount ||
		positiveTrainCount == 0 ||
		positiveValCount == 0 {
		trainingWriteError(
			w,
			http.StatusBadRequest,
			fmt.Sprintf(
				"Zu wenige YOLO26-Beispiele. Train=%d (%d positiv), Val=%d (%d positiv). Benötigt: mindestens %d Train, %d Val und je ein positives Beispiel.",
				trainCount,
				positiveTrainCount,
				valCount,
				positiveValCount,
				minDetectorTrainCount,
				minDetectorValCount,
			),
		)
		return
	}

	if !fileExistsNonEmpty(poseDatasetYAML) ||
		poseTrainCount < minPoseTrainCount ||
		poseValCount < minPoseValCount {
		trainingWriteError(
			w,
			http.StatusBadRequest,
			fmt.Sprintf(
				"Zu wenige YOLO26-Pose-Beispiele. Train=%d, Val=%d. Benoetigt: mindestens %d Train und %d Val.",
				poseTrainCount,
				poseValCount,
				minPoseTrainCount,
				minPoseValCount,
			),
		)
		return
	}

startTraining:
	ctx, cancel := context.WithCancel(context.Background())

	trainingStartJob(cancel)

	go trainingRunJob(ctx, root, feedbackCount)

	trainingWriteJSON(w, http.StatusAccepted, map[string]any{
		"ok":       true,
		"message":  "Training gestartet.",
		"training": trainingGetJobStatus(),
		"detector": map[string]any{
			"trainCount":         trainCount,
			"valCount":           valCount,
			"positiveTrainCount": positiveTrainCount,
			"positiveValCount":   positiveValCount,
			"requiredTrain":      minDetectorTrainCount,
			"requiredVal":        minDetectorValCount,
			"datasetYAML":        detectorDatasetYAML,
			"usesSceneCLIP":      false,
			"usesSceneKNN":       false,
			"source":             "yolo26_detector",
			"detectsPosition":    false,
		},
		"pose": map[string]any{
			"trainCount":    poseTrainCount,
			"valCount":      poseValCount,
			"requiredTrain": minPoseTrainCount,
			"requiredVal":   minPoseValCount,
			"datasetYAML":   poseDatasetYAML,
			"source":        "yolo26_pose",
		},
		"videomae": map[string]any{
			"eligibleCount": videoMAEEligibleCount,
			"trainCount":    videoMAETrainCount,
			"valCount":      videoMAEValCount,
			"requiredTrain": minVideoMAETrainCount,
			"requiredVal":   minVideoMAEValCount,
			"manifest":      videoMAEManifest,
			"source":        "videomae_clip",
		},
	})
}

func trainingCancelHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodPost && r.Method != http.MethodDelete {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	trainingJob.mu.Lock()
	status := trainingJob.status
	cancel := trainingJob.cancel
	trainingJob.mu.Unlock()

	if !status.Running {
		trainingWriteJSON(w, http.StatusOK, map[string]any{
			"ok":       true,
			"message":  "Es läuft kein Training.",
			"training": status,
		})
		return
	}

	trainingSetJobStatus(func(s *TrainingJobStatus) {
		s.Step = "Training wird abgebrochen…"
		s.Message = ""
		s.Error = ""
	})

	if cancel != nil {
		cancel()
	}

	trainingWriteJSON(w, http.StatusAccepted, map[string]any{
		"ok":       true,
		"message":  "Training wird abgebrochen.",
		"training": trainingGetJobStatus(),
	})
}

func trainingRunJob(ctx context.Context, root string, count int) {
	python := trainingPythonExe()

	cleanOutput := func(text string) string {
		out := strings.TrimSpace(text)
		if len(out) > 1500 {
			out = out[:1500] + "…"
		}
		return out
	}

	trainingSetJobStatus(func(s *TrainingJobStatus) {
		s.Progress = 10
		s.Step = "YOLO26-Daten werden geprüft…"
	})

	detectorOutput := ""
	detectorStatus := "skipped"
	poseOutput := ""
	poseStatus := "skipped"
	videoMAEOutput := ""
	videoMAEStatus := "skipped"

	detectorDatasetYAML := filepath.Join(root, "detector", "dataset", "dataset.yaml")
	detectorTrainImages := filepath.Join(root, "detector", "dataset", "images", "train")
	detectorTrainLabels := filepath.Join(root, "detector", "dataset", "labels", "train")
	detectorValImages := filepath.Join(root, "detector", "dataset", "images", "val")
	detectorValLabels := filepath.Join(root, "detector", "dataset", "labels", "val")

	if err := trainingEnsureDetectorValidationSample(root); err != nil {
		appLogln("⚠️ detector val sample ensure failed:", err)
	}

	trainCount := trainingCountDetectorSamples(detectorTrainImages, detectorTrainLabels)
	valCount := trainingCountDetectorSamples(detectorValImages, detectorValLabels)
	positiveTrainCount := trainingCountPositiveDetectorSamples(detectorTrainImages, detectorTrainLabels)
	positiveValCount := trainingCountPositiveDetectorSamples(detectorValImages, detectorValLabels)

	fmt.Printf(
		"🔎 detector data: train=%d (%d positive) val=%d (%d positive) yaml=%v\n",
		trainCount,
		positiveTrainCount,
		valCount,
		positiveValCount,
		fileExistsNonEmpty(detectorDatasetYAML),
	)

	if fileExistsNonEmpty(detectorDatasetYAML) &&
		trainCount >= minDetectorTrainCount &&
		valCount >= minDetectorValCount &&
		positiveTrainCount > 0 &&
		positiveValCount > 0 {
		trainingSetJobStatus(func(s *TrainingJobStatus) {
			s.Progress = 15
			s.Step = "YOLO26 Detector wird trainiert…"
		})

		detectorScript := trainingScriptPath("train_detector_model.py")
		detectorOut, detectorErr := trainingRunCommandStreaming(
			ctx,
			python,
			detectorScript,
			func(line string) bool {
				return trainingHandleProgressLine(
					line,
					15,
					58,
					"YOLO26 Detector wird trainiert…",
				)
			},
			"--root", root,
			"--base", "yolo26n.pt",
			"--epochs", strconv.Itoa(trainingDetectorEpochs()),
			"--imgsz", "640",
		)

		if errors.Is(detectorErr, errTrainingCancelled) {
			appLogln("⛔ YOLO26 detector training cancelled")
			trainingFinishCancelled(root)
			return
		}

		detectorOutput = detectorOut
		detectorOutputClean := cleanOutput(detectorOutput)

		if detectorErr != nil {
			detectorStatus = "failed"

			appLogln("⚠️ YOLO26 detector training failed:", detectorErr)
			if detectorOutputClean != "" {
				appLogln("⚠️ YOLO26 detector output:", detectorOutputClean)
			}
		} else {
			detectorStatus = "trained"

			if detectorOutputClean != "" {
				appLogln("✅ YOLO26 detector training:", detectorOutputClean)
			}
		}
	} else {
		detectorStatus = "skipped_no_detector_data"
		detectorOutput = fmt.Sprintf(
			"YOLO26 Detector übersprungen: zu wenige Beispiele. Train=%d (%d positiv), Val=%d (%d positiv). Benötigt: mindestens %d Train, %d Val und je ein positives Beispiel.",
			trainCount,
			positiveTrainCount,
			valCount,
			positiveValCount,
			minDetectorTrainCount,
			minDetectorValCount,
		)

		appLogln("⚠️", detectorOutput)
	}

	detectorOutputClean := cleanOutput(detectorOutput)
	poseDatasetYAML := filepath.Join(root, "pose", "dataset", "dataset.yaml")
	poseTrainImages := filepath.Join(root, "pose", "dataset", "images", "train")
	poseTrainLabels := filepath.Join(root, "pose", "dataset", "labels", "train")
	poseValImages := filepath.Join(root, "pose", "dataset", "images", "val")
	poseValLabels := filepath.Join(root, "pose", "dataset", "labels", "val")

	trainingSetJobStatus(func(s *TrainingJobStatus) {
		if s.Progress < 60 {
			s.Progress = 60
		}
		s.Step = "YOLO26 Pose-Daten werden aufgebaut..."
	})

	if written, err := trainingSyncPoseDataset(root); err != nil {
		poseStatus = "failed"
		poseOutput = "YOLO26 Pose-Dataset konnte nicht aufgebaut werden: " + err.Error()
		appLogln(poseOutput)
	} else {
		appLogln("pose samples synced:", written)
	}

	if err := trainingEnsurePoseValidationSample(root); err != nil {
		appLogln("pose val sample ensure failed:", err)
	}

	poseTrainCount := trainingCountDetectorSamples(poseTrainImages, poseTrainLabels)
	poseValCount := trainingCountDetectorSamples(poseValImages, poseValLabels)

	fmt.Printf(
		"pose data: train=%d val=%d yaml=%v\n",
		poseTrainCount,
		poseValCount,
		fileExistsNonEmpty(poseDatasetYAML),
	)

	if poseStatus != "failed" &&
		fileExistsNonEmpty(poseDatasetYAML) &&
		poseTrainCount >= minPoseTrainCount &&
		poseValCount >= minPoseValCount {
		trainingSetJobStatus(func(s *TrainingJobStatus) {
			if s.Progress < 62 {
				s.Progress = 62
			}
			s.Step = "YOLO26 Pose wird trainiert..."
		})

		poseBasePath := "yolo26n-pose.pt"
		if p, err := embeddedPoseModelPath(); err == nil && fileExistsNonEmpty(p) {
			poseBasePath = p
		}

		poseScript := trainingScriptPath("train_pose_model.py")
		poseOut, poseErr := trainingRunCommandStreaming(
			ctx,
			python,
			poseScript,
			func(line string) bool {
				return trainingHandleProgressLine(
					line,
					62,
					82,
					"YOLO26 Pose wird trainiert...",
				)
			},
			"--root", root,
			"--base", poseBasePath,
			"--epochs", strconv.Itoa(trainingDetectorEpochs()),
			"--imgsz", "640",
		)

		if errors.Is(poseErr, errTrainingCancelled) {
			appLogln("YOLO26 pose training cancelled")
			trainingFinishCancelled(root)
			return
		}

		poseOutput = poseOut
		poseOutputClean := cleanOutput(poseOutput)

		if poseErr != nil {
			poseStatus = "failed"

			appLogln("YOLO26 pose training failed:", poseErr)
			if poseOutputClean != "" {
				appLogln("YOLO26 pose output:", poseOutputClean)
			}
		} else {
			poseStatus = "trained"

			if poseOutputClean != "" {
				appLogln("YOLO26 pose training:", poseOutputClean)
			}
		}
	} else if poseStatus != "failed" {
		poseStatus = "skipped_no_pose_data"
		poseOutput = fmt.Sprintf(
			"YOLO26 Pose uebersprungen: zu wenige Skeleton-Beispiele. Train=%d, Val=%d. Benoetigt: mindestens %d Train und %d Val.",
			poseTrainCount,
			poseValCount,
			minPoseTrainCount,
			minPoseValCount,
		)
		appLogln(poseOutput)
	}

	poseOutputClean := cleanOutput(poseOutput)

	trainingSetJobStatus(func(s *TrainingJobStatus) {
		if s.Progress < 84 {
			s.Progress = 84
		}
		s.Step = "VideoMAE Clip-Daten werden aufgebaut..."
	})

	videoMAETrainCount, videoMAEValCount, videoMAEWritten, videoMAESyncErr :=
		trainingSyncVideoMAEDataset(ctx, root)
	if errors.Is(videoMAESyncErr, context.Canceled) || errors.Is(videoMAESyncErr, errTrainingCancelled) {
		appLogln("VideoMAE dataset sync cancelled")
		trainingFinishCancelled(root)
		return
	}
	if videoMAESyncErr != nil {
		videoMAEStatus = "failed"
		videoMAEOutput = "VideoMAE-Dataset konnte nicht aufgebaut werden: " + videoMAESyncErr.Error()
		appLogln(videoMAEOutput)
	} else {
		appLogf(
			"VideoMAE samples synced: written=%d train=%d val=%d",
			videoMAEWritten,
			videoMAETrainCount,
			videoMAEValCount,
		)
	}

	if videoMAEStatus != "failed" &&
		videoMAETrainCount >= minVideoMAETrainCount &&
		videoMAEValCount >= minVideoMAEValCount {
		trainingSetJobStatus(func(s *TrainingJobStatus) {
			if s.Progress < 86 {
				s.Progress = 86
			}
			s.Step = "VideoMAE Clip-Classifier wird trainiert..."
		})

		videoMAEScript := trainingScriptPath("train_videomae_model.py")
		videoMAEArgs := []string{
			"--root", root,
			"--epochs", strconv.Itoa(trainingVideoMAEEpochs()),
			"--batch-size", strconv.Itoa(trainingVideoMAEBatchSize()),
			"--num-frames", strconv.Itoa(trainingVideoMAENumFrames),
		}
		if base := strings.TrimSpace(os.Getenv("VIDEOMAE_BASE_MODEL")); base != "" {
			videoMAEArgs = append(videoMAEArgs, "--base", base)
		}

		videoMAEOut, videoMAEErr := trainingRunCommandStreaming(
			ctx,
			python,
			videoMAEScript,
			func(line string) bool {
				return trainingHandleProgressLine(
					line,
					86,
					98,
					"VideoMAE Clip-Classifier wird trainiert...",
				)
			},
			videoMAEArgs...,
		)

		if errors.Is(videoMAEErr, errTrainingCancelled) {
			appLogln("VideoMAE training cancelled")
			trainingFinishCancelled(root)
			return
		}

		videoMAEOutput = videoMAEOut
		videoMAEOutputClean := cleanOutput(videoMAEOutput)

		if videoMAEErr != nil {
			videoMAEStatus = "failed"
			appLogln("VideoMAE training failed:", videoMAEErr)
			if videoMAEOutputClean != "" {
				appLogln("VideoMAE output:", videoMAEOutputClean)
			}
		} else {
			videoMAEStatus = "trained"
			if videoMAEOutputClean != "" {
				appLogln("VideoMAE training:", videoMAEOutputClean)
			}
		}
	} else if videoMAEStatus != "failed" {
		videoMAEStatus = "skipped_no_videomae_data"
		videoMAEOutput = fmt.Sprintf(
			"VideoMAE uebersprungen: zu wenige Clip-Beispiele. Train=%d, Val=%d. Benoetigt: mindestens %d Train und %d Val.",
			videoMAETrainCount,
			videoMAEValCount,
			minVideoMAETrainCount,
			minVideoMAEValCount,
		)
		appLogln(videoMAEOutput)
	}

	videoMAEOutputClean := cleanOutput(videoMAEOutput)

	message := "Training abgeschlossen."
	errorText := ""

	switch detectorStatus {
	case "trained":
		message = "Training abgeschlossen. YOLO26 Detector wurde trainiert."

	case "skipped_no_detector_data":
		message = detectorOutput

	case "failed":
		message = "YOLO26 Detector ist fehlgeschlagen."
		if detectorOutputClean != "" {
			message += " Grund: " + detectorOutputClean
		}
		errorText = message

	default:
		message = "Training abgeschlossen, aber YOLO26 wurde nicht trainiert."
		if detectorOutputClean != "" {
			message += " Ausgabe: " + detectorOutputClean
		}
	}

	if poseStatus == "trained" && detectorStatus == "trained" {
		message = "Training abgeschlossen. YOLO26 Detector und YOLO26 Pose wurden trainiert."
	} else if poseStatus == "trained" && detectorStatus != "failed" {
		message = "Training abgeschlossen. YOLO26 Pose wurde trainiert."
	} else if poseStatus == "skipped_no_pose_data" && detectorStatus == "trained" {
		message += " YOLO26 Pose wurde uebersprungen: zu wenige Skeleton-Beispiele."
	} else if poseStatus == "failed" {
		if detectorStatus == "failed" {
			message += " YOLO26 Pose ist ebenfalls fehlgeschlagen."
		} else {
			message = "YOLO26 Pose ist fehlgeschlagen."
		}
		if poseOutputClean != "" {
			message += " Grund: " + poseOutputClean
		}
		errorText = message
	}

	if videoMAEStatus == "trained" {
		if !strings.Contains(message, "Training abgeschlossen.") {
			message = "Training abgeschlossen. " + message
		}
		message += " VideoMAE wurde trainiert."
	} else if videoMAEStatus == "skipped_no_videomae_data" {
		if detectorStatus == "trained" || poseStatus == "trained" {
			message += " VideoMAE wurde uebersprungen: zu wenige Clip-Beispiele."
		}
	} else if videoMAEStatus == "failed" {
		message += " VideoMAE ist fehlgeschlagen."
		if videoMAEOutputClean != "" {
			message += " Grund: " + videoMAEOutputClean
		}
	}

	if detectorStatus == "trained" || poseStatus == "trained" || videoMAEStatus == "trained" {
		// Verlaufseintrag schreiben, solange die Job-Startzeit für die Dauer noch verfügbar ist.
		trainingAppendRunHistory(root)
	}

	trainingSetJobStatus(func(s *TrainingJobStatus) {
		finishedAt := time.Now().UTC()

		var durationMs int64
		if startedAt, err := time.Parse(time.RFC3339, strings.TrimSpace(s.StartedAt)); err == nil {
			durationMs = finishedAt.Sub(startedAt).Milliseconds()
			if durationMs < 0 {
				durationMs = 0
			}
		}

		s.Running = false
		s.Progress = 100
		s.Step = "Training abgeschlossen."
		s.Message = message
		s.Error = errorText
		s.FinishedAt = finishedAt.Format(time.RFC3339)
		s.DurationMs = durationMs
		s.PreviewURL = ""
	})

	trainingClearJobCancel()
}

func trainingStatsHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodGet {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	stats, err := trainingBuildStats(root)
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	trainingWriteJSON(w, http.StatusOK, stats)
}

func trainingBuildStats(root string) (*TrainingStatsResponse, error) {
	grouped, err := trainingGroupedLabels()
	if err != nil {
		// Fallback: Stats sollen trotzdem funktionieren, auch wenn Label-Gruppierung scheitert.
		fallbackLabels := defaultTrainingLabelsFromJSON()

		grouped = TrainingGroupedLabels{
			People:       fallbackLabels.People,
			SexPositions: fallbackLabels.SexPositions,
			BodyParts:    fallbackLabels.BodyParts,
			Objects:      fallbackLabels.Objects,
			Clothing:     fallbackLabels.Clothing,
		}
	}

	peopleSet := stringSet(grouped.People)
	sexPositionSet := stringSet(grouped.SexPositions)
	bodyPartSet := stringSet(grouped.BodyParts)
	objectSet := stringSet(grouped.Objects)
	clothingSet := stringSet(grouped.Clothing)

	peopleCounts := map[string]int{}
	sexPositionCounts := map[string]int{}
	bodyPartCounts := map[string]int{}
	objectCounts := map[string]int{}
	clothingCounts := map[string]int{}

	stats := &TrainingStatsResponse{
		OK: true,
		Labels: TrainingStatsLabels{
			People:       []TrainingLabelStat{},
			SexPositions: []TrainingLabelStat{},
			BodyParts:    []TrainingLabelStat{},
			Objects:      []TrainingLabelStat{},
			Clothing:     []TrainingLabelStat{},
		},
	}

	feedbackPath := filepath.Join(root, "feedback.jsonl")
	b, err := os.ReadFile(feedbackPath)
	if err != nil {
		if os.IsNotExist(err) {
			stats.SampleCount = trainingCountSampleFiles(filepath.Join(root, "samples"))
			trainingApplyStatsModelInfo(root, stats)
			return stats, nil
		}

		return nil, err
	}

	for _, line := range strings.Split(string(b), "\n") {
		line = strings.TrimSpace(line)
		if line == "" {
			continue
		}

		var annotation TrainingAnnotation
		if err := json.Unmarshal([]byte(line), &annotation); err != nil {
			continue
		}

		stats.FeedbackCount++

		if annotation.Negative {
			stats.NegativeCount++
		} else if annotation.Accepted {
			stats.AcceptedCount++
		} else {
			stats.CorrectedCount++
		}

		effective := trainingEffectiveCorrection(annotation)

		sexPosition := normalizeSexPositionLabel(effective.SexPosition)
		if len(sexPositionSet) == 0 || sexPositionSet[sexPosition] {
			sexPositionCounts[sexPosition]++
		}

		for _, label := range effective.BodyPartsPresent {
			clean := strings.TrimSpace(label)
			if clean == "" {
				continue
			}
			if len(bodyPartSet) == 0 || bodyPartSet[clean] {
				bodyPartCounts[clean]++
			}
		}

		for _, label := range effective.ObjectsPresent {
			clean := strings.TrimSpace(label)
			if clean == "" {
				continue
			}
			if len(objectSet) == 0 || objectSet[clean] {
				objectCounts[clean]++
			}
		}

		for _, label := range effective.ClothingPresent {
			clean := strings.TrimSpace(label)
			if clean == "" {
				continue
			}
			if len(clothingSet) == 0 || clothingSet[clean] {
				clothingCounts[clean]++
			}
		}

		for _, box := range effective.Boxes {
			label := strings.TrimSpace(box.Label)
			if label == "" {
				continue
			}

			stats.BoxCount++

			switch {
			case peopleSet[label]:
				peopleCounts[label]++

			case bodyPartSet[label]:
				bodyPartCounts[label]++

			case objectSet[label]:
				objectCounts[label]++

			case clothingSet[label]:
				clothingCounts[label]++
			}
		}
	}

	stats.SampleCount = trainingCountSampleFiles(filepath.Join(root, "samples"))
	trainingApplyStatsModelInfo(root, stats)

	stats.Labels = TrainingStatsLabels{
		// Personen/Box-Labels brauchen mehr Beispiele, weil der Detector Boxen lernen muss.
		People: trainingStatsMapToList(peopleCounts, 20),

		// Scene-Positionen sind Sample-Labels, hier reichen grob weniger pro Klasse.
		SexPositions: trainingStatsMapToList(sexPositionCounts, 8),

		// Detector-Klassen: grob 15 Beispiele pro Label als solide Untergrenze.
		BodyParts: trainingStatsMapToList(bodyPartCounts, 15),
		Objects:   trainingStatsMapToList(objectCounts, 15),
		Clothing:  trainingStatsMapToList(clothingCounts, 15),
	}

	stats.Confidence = trainingOverallConfidence(
		stats.FeedbackCount,
		stats.BoxCount,
		stats.AcceptedCount,
		stats.CorrectedCount,
		stats.Labels,
	)

	return stats, nil
}

func trainingEffectiveCorrection(annotation TrainingAnnotation) TrainingCorrection {
	if annotation.Negative {
		return *trainingNegativeCorrection()
	}

	if annotation.Correction != nil {
		return *annotation.Correction
	}

	p := annotation.Prediction

	return TrainingCorrection{
		SexPosition:      p.SexPosition,
		PeoplePresent:    trainingScoredLabelsToStrings(p.PeoplePresent),
		BodyPartsPresent: trainingScoredLabelsToStrings(p.BodyPartsPresent),
		ObjectsPresent:   trainingScoredLabelsToStrings(p.ObjectsPresent),
		ClothingPresent:  trainingScoredLabelsToStrings(p.ClothingPresent),
		Boxes:            p.Boxes,
	}
}

func trainingScoredLabelsToStrings(values []TrainingScoredLabel) []string {
	out := make([]string, 0, len(values))
	seen := map[string]bool{}

	for _, value := range values {
		label := strings.TrimSpace(value.Label)
		if label == "" || seen[label] {
			continue
		}

		seen[label] = true
		out = append(out, label)
	}

	return out
}

func trainingStatsMapToList(values map[string]int, target int) []TrainingLabelStat {
	out := make([]TrainingLabelStat, 0, len(values))

	for label, count := range values {
		label = strings.TrimSpace(label)
		if label == "" || count <= 0 {
			continue
		}

		out = append(out, TrainingLabelStat{
			Label:      label,
			Count:      count,
			Confidence: trainingLabelConfidence(count, target),
		})
	}

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

		return out[i].Count > out[j].Count
	})

	return out
}

func trainingCountSampleFiles(samplesDir string) int {
	entries, err := os.ReadDir(samplesDir)
	if err != nil {
		return 0
	}

	count := 0

	for _, entry := range entries {
		if entry.IsDir() {
			continue
		}

		if strings.ToLower(filepath.Ext(entry.Name())) == ".json" {
			count++
		}
	}

	return count
}

func trainingStatusHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodGet {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	job := trainingGetJobStatus()

	if !job.Running {
		if err := trainingEnsureDetectorDirs(root); err != nil {
			trainingWriteError(w, http.StatusInternalServerError, err.Error())
			return
		}
		if err := trainingEnsurePoseDirs(root); err != nil {
			trainingWriteError(w, http.StatusInternalServerError, err.Error())
			return
		}
		if err := trainingEnsureVideoMAEDirs(root); err != nil {
			trainingWriteError(w, http.StatusInternalServerError, err.Error())
			return
		}

		if err := trainingEnsureDetectorValidationSample(root); err != nil {
			appLogln("⚠️ detector val sample ensure failed:", err)
		}
		if err := trainingEnsurePoseValidationSample(root); err != nil {
			appLogln("pose val sample ensure failed:", err)
		}
	}

	feedbackPath := filepath.Join(root, "feedback.jsonl")
	feedbackCount, _ := trainingCountAnnotations(feedbackPath)

	detectorDatasetYAML := filepath.Join(root, "detector", "dataset", "dataset.yaml")
	detectorTrainImages := filepath.Join(root, "detector", "dataset", "images", "train")
	detectorTrainLabels := filepath.Join(root, "detector", "dataset", "labels", "train")
	detectorValImages := filepath.Join(root, "detector", "dataset", "images", "val")
	detectorValLabels := filepath.Join(root, "detector", "dataset", "labels", "val")
	poseDatasetYAML := filepath.Join(root, "pose", "dataset", "dataset.yaml")
	poseTrainImages := filepath.Join(root, "pose", "dataset", "images", "train")
	poseTrainLabels := filepath.Join(root, "pose", "dataset", "labels", "train")
	poseValImages := filepath.Join(root, "pose", "dataset", "images", "val")
	poseValLabels := filepath.Join(root, "pose", "dataset", "labels", "val")
	detectorModel := trainingResolveDetectorModel(root)
	poseModel := trainingResolvePoseModel(root)
	videoMAEModel := trainingResolveVideoMAEModel(root)
	videoMAEManifest := trainingVideoMAEManifestPath(root)

	trainCount := trainingCountDetectorSamples(detectorTrainImages, detectorTrainLabels)
	valCount := trainingCountDetectorSamples(detectorValImages, detectorValLabels)
	positiveTrainCount := trainingCountPositiveDetectorSamples(detectorTrainImages, detectorTrainLabels)
	positiveValCount := trainingCountPositiveDetectorSamples(detectorValImages, detectorValLabels)
	poseTrainCount := trainingCountDetectorSamples(poseTrainImages, poseTrainLabels)
	poseValCount := trainingCountDetectorSamples(poseValImages, poseValLabels)
	videoMAETrainCount, videoMAEValCount := trainingCountVideoMAEManifestSamples(root)
	videoMAEEligibleCount, _ := trainingCountVideoMAEEligibleAnnotations(root)

	datasetReady := fileExistsNonEmpty(detectorDatasetYAML)
	detectorDataReady := datasetReady &&
		trainCount >= minDetectorTrainCount &&
		valCount >= minDetectorValCount &&
		positiveTrainCount > 0 &&
		positiveValCount > 0
	poseDatasetReady := fileExistsNonEmpty(poseDatasetYAML)
	poseDataReady := poseDatasetReady &&
		poseTrainCount >= minPoseTrainCount &&
		poseValCount >= minPoseValCount
	videoMAEDatasetReady := fileExistsNonEmpty(videoMAEManifest)
	videoMAEDataReady := (videoMAETrainCount >= minVideoMAETrainCount &&
		videoMAEValCount >= minVideoMAEValCount) ||
		videoMAEEligibleCount >= minVideoMAETrainCount

	canTrain := feedbackCount >= minTrainingFeedbackCount &&
		(detectorDataReady || poseDataReady || videoMAEDataReady)

	trainingWriteJSON(w, http.StatusOK, map[string]any{
		"ok":            true,
		"feedbackCount": feedbackCount,
		"requiredCount": minTrainingFeedbackCount,
		"canTrain":      canTrain,

		"training": job,

		"detector": map[string]any{
			"source":          "yolo26_detector",
			"usesSceneCLIP":   false,
			"usesSceneKNN":    false,
			"usesResNet18KNN": false,

			"detectsPeople":      true,
			"detectsGender":      true,
			"detectsSexPosition": false,
			"detectsBodyParts":   true,
			"detectsObjects":     true,
			"detectsClothing":    true,
			"detectsBoxes":       true,

			"trainCount":         trainCount,
			"valCount":           valCount,
			"positiveTrainCount": positiveTrainCount,
			"positiveValCount":   positiveValCount,
			"requiredTrain":      minDetectorTrainCount,
			"requiredVal":        minDetectorValCount,

			"datasetReady": datasetReady,
			"datasetYAML":  detectorDatasetYAML,
			"dataReady":    detectorDataReady,

			"modelExists":        detectorModel.EffectiveExists,
			"modelPath":          detectorModel.EffectivePath,
			"trainedModelExists": detectorModel.TrainedExists,
			"trainedModelPath":   detectorModel.BestPath,
			"modelSource":        detectorModel.Source,
		},

		"pose": map[string]any{
			"source":              "yolo26_pose",
			"usesKeypoints":       true,
			"predictsPersons":     true,
			"predictsSexPosition": poseModel.TrainedExists,
			"trainedFromFeedback": true,
			"trainCount":          poseTrainCount,
			"valCount":            poseValCount,
			"requiredTrain":       minPoseTrainCount,
			"requiredVal":         minPoseValCount,
			"datasetReady":        poseDatasetReady,
			"datasetYAML":         poseDatasetYAML,
			"dataReady":           poseDataReady,
			"modelExists":         poseModel.EffectiveExists,
			"modelPath":           poseModel.EffectivePath,
			"trainedModelExists":  poseModel.TrainedExists,
			"trainedModelPath":    poseModel.BestPath,
			"modelSource":         poseModel.Source,
		},

		"scene": map[string]any{
			"source":                 "videomae_clip",
			"usesVideoMAE":           true,
			"usesSceneCLIP":          false,
			"usesSceneKNN":           false,
			"usesResNet18KNN":        false,
			"usesLogisticRegression": false,

			"predictsSexPosition": videoMAEModel.TrainedExists,
			"predictsPeople":      false,
			"predictsGender":      false,
			"predictsBodyParts":   false,
			"predictsObjects":     false,
			"predictsClothing":    false,
			"predictsBoxes":       false,

			"feedbackCount": feedbackCount,
			"eligibleCount": videoMAEEligibleCount,
			"trainCount":    videoMAETrainCount,
			"valCount":      videoMAEValCount,
			"requiredTrain": minVideoMAETrainCount,
			"requiredVal":   minVideoMAEValCount,
			"requiredCount": minVideoMAETrainCount,
			"datasetReady":  videoMAEDatasetReady,
			"manifest":      videoMAEManifest,
			"dataReady":     videoMAEDataReady,
			"modelReady":    videoMAEModel.EffectiveExists,
			"modelExists":   videoMAEModel.EffectiveExists,
			"modelPath":     videoMAEModel.EffectivePath,
			"modelSource":   videoMAEModel.Source,
		},

		"pipeline": map[string]any{
			"variant": "YOLO26_VIDEO_CLIP_HYBRID",

			"peopleSource":      "yolo26_detector",
			"genderSource":      "yolo26_detector",
			"sexPositionSource": "yolo26_pose+box_context+videomae_clip",
			"bodyPartsSource":   "yolo26_detector",
			"objectsSource":     "yolo26_detector",
			"clothingSource":    "yolo26_detector",
			"boxesSource":       "yolo26_detector",

			"usesSceneKNNForDetection": false,
			"usesSceneCLIP":            false,
			"usesSceneKNN":             false,
			"usesVideoMAE":             true,
			"usesYOLOForDetection":     true,
			"usesYOLOForSexPosition":   true,
		},
	})
}

func trainingApplyStatsModelInfo(root string, stats *TrainingStatsResponse) {
	detectorAvailable := trainingStatsModelAvailableFor(root, "detector")
	detectorInfo := trainingReadModelInfoFor(root, "detector")
	poseAvailable := trainingStatsModelAvailableFor(root, "pose")
	poseInfo := trainingReadModelInfoFor(root, "pose")
	videoMAEModel := trainingResolveVideoMAEModel(root)
	videoMAEInfo := trainingReadModelInfoFor(root, "videomae")

	// modelAvailable/modelInfo bleiben aus Kompatibilitaetsgruenden der Detector.
	stats.ModelAvailable = detectorAvailable
	stats.ModelInfo = detectorInfo
	stats.DetectorModelAvailable = detectorAvailable
	stats.DetectorModelInfo = detectorInfo
	stats.PoseModelAvailable = poseAvailable
	stats.PoseModelInfo = poseInfo
	stats.VideoMAEModelAvailable = videoMAEModel.EffectiveExists
	stats.VideoMAEModelInfo = videoMAEInfo
}

func trainingStatsModelAvailable(root string) bool {
	return trainingStatsModelAvailableFor(root, "detector")
}

func trainingStatsModelAvailableFor(root string, kind string) bool {
	modelPath := filepath.Join(root, kind, "model", "best.pt")
	if kind == "videomae" {
		modelPath = filepath.Join(root, kind, "model", "config.json")
	}
	return fileExistsNonEmpty(modelPath)
}

// trainingReadModelInfo liest Versions-/Datums-Infos zum aktuell trainierten
// Detector-Modell. Datum/Version stammen primär aus status.json (vom Trainingsskript),
// Fallback ist die Änderungszeit der best.pt-Datei.
func trainingReadModelInfo(root string) *TrainingModelInfo {
	return trainingReadModelInfoFor(root, "detector")
}

func trainingReadModelInfoFor(root string, kind string) *TrainingModelInfo {
	modelPath := filepath.Join(root, kind, "model", "best.pt")
	if kind == "videomae" {
		modelPath = filepath.Join(root, kind, "model", "config.json")
	}

	fi, err := os.Stat(modelPath)
	if err != nil || fi.IsDir() || fi.Size() <= 0 {
		return nil
	}

	info := &TrainingModelInfo{
		TrainedAt:   fi.ModTime().UTC().Format(time.RFC3339),
		TrainedAtMs: fi.ModTime().UnixMilli(),
	}

	statusPath := filepath.Join(root, kind, "model", "status.json")
	if b, err := os.ReadFile(statusPath); err == nil {
		var raw struct {
			TrainedAt    string  `json:"trainedAt"`
			Epochs       int     `json:"epochs"`
			TrainSamples int     `json:"trainSamples"`
			ValSamples   int     `json:"valSamples"`
			Imgsz        int     `json:"imgsz"`
			Device       string  `json:"device"`
			MAP50        float64 `json:"mAP50"`
			MAP5095      float64 `json:"mAP5095"`
		}

		if json.Unmarshal(b, &raw) == nil {
			if trimmed := strings.TrimSpace(raw.TrainedAt); trimmed != "" {
				if t, err := time.Parse(time.RFC3339, trimmed); err == nil {
					info.TrainedAt = t.UTC().Format(time.RFC3339)
					info.TrainedAtMs = t.UnixMilli()
				}
			}

			info.Epochs = raw.Epochs
			info.TrainSamples = raw.TrainSamples
			info.ValSamples = raw.ValSamples
			info.Imgsz = raw.Imgsz
			info.Device = strings.TrimSpace(raw.Device)
			info.MAP50 = raw.MAP50
			info.MAP5095 = raw.MAP5095
		}
	}

	return info
}

type TrainingHistoryEntry struct {
	TrainedAt    string  `json:"trainedAt,omitempty"`
	TrainedAtMs  int64   `json:"trainedAtMs,omitempty"`
	DurationMs   int64   `json:"durationMs,omitempty"`
	Epochs       int     `json:"epochs,omitempty"`
	TrainSamples int     `json:"trainSamples,omitempty"`
	ValSamples   int     `json:"valSamples,omitempty"`
	Imgsz        int     `json:"imgsz,omitempty"`
	Device       string  `json:"device,omitempty"`
	MAP50        float64 `json:"map50,omitempty"`
	MAP5095      float64 `json:"map5095,omitempty"`
}

type TrainingHistoryResponse struct {
	OK      bool                   `json:"ok"`
	Entries []TrainingHistoryEntry `json:"entries"`
}

func trainingHistoryPath(root string) string {
	return filepath.Join(root, "detector", "training_history.jsonl")
}

// trainingAppendRunHistory hängt nach einem erfolgreichen Trainingslauf einen
// Verlaufseintrag an (Datum, mAP, Samples, Epochen, Dauer).
func trainingAppendRunHistory(root string) {
	info := trainingReadModelInfo(root)
	if info == nil {
		return
	}

	entry := TrainingHistoryEntry{
		TrainedAt:    info.TrainedAt,
		TrainedAtMs:  info.TrainedAtMs,
		Epochs:       info.Epochs,
		TrainSamples: info.TrainSamples,
		ValSamples:   info.ValSamples,
		Imgsz:        info.Imgsz,
		Device:       info.Device,
		MAP50:        info.MAP50,
		MAP5095:      info.MAP5095,
	}

	// Dauer aus der Startzeit des aktuellen Jobs ableiten.
	job := trainingGetJobStatus()
	if startedAt, err := time.Parse(time.RFC3339, strings.TrimSpace(job.StartedAt)); err == nil {
		if ms := time.Now().UTC().Sub(startedAt).Milliseconds(); ms > 0 {
			entry.DurationMs = ms
		}
	}

	b, err := json.Marshal(entry)
	if err != nil {
		return
	}

	path := trainingHistoryPath(root)
	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
		appLogln("⚠️ training history dir konnte nicht erstellt werden:", err)
		return
	}

	f, err := os.OpenFile(path, os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0o644)
	if err != nil {
		appLogln("⚠️ training history konnte nicht geöffnet werden:", err)
		return
	}
	defer f.Close()

	if _, err := f.Write(append(b, '\n')); err != nil {
		appLogln("⚠️ training history konnte nicht geschrieben werden:", err)
	}
}

// trainingReadHistory liest die Verlaufseinträge, neueste zuerst.
func trainingReadHistory(root string, limit int) []TrainingHistoryEntry {
	b, err := os.ReadFile(trainingHistoryPath(root))
	if err != nil {
		return nil
	}

	lines := strings.Split(string(b), "\n")
	entries := make([]TrainingHistoryEntry, 0, len(lines))

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

		var e TrainingHistoryEntry
		if json.Unmarshal([]byte(line), &e) != nil {
			continue
		}

		entries = append(entries, e)
	}

	// Datei ist chronologisch → umdrehen, damit der neueste Lauf oben steht.
	for i, j := 0, len(entries)-1; i < j; i, j = i+1, j-1 {
		entries[i], entries[j] = entries[j], entries[i]
	}

	if limit > 0 && len(entries) > limit {
		entries = entries[:limit]
	}

	return entries
}

func trainingHistoryHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodGet {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	entries := trainingReadHistory(root, 50)
	if entries == nil {
		entries = []TrainingHistoryEntry{}
	}

	trainingWriteJSON(w, http.StatusOK, TrainingHistoryResponse{
		OK:      true,
		Entries: entries,
	})
}

func trainingConfidenceFromScore(score float64) TrainingConfidence {
	if math.IsNaN(score) || math.IsInf(score, 0) {
		score = 0
	}

	score = clamp01(score)

	level := "none"
	label := "Keine"

	switch {
	case score >= 0.75:
		level = "high"
		label = "Hoch"
	case score >= 0.45:
		level = "mid"
		label = "Mittel"
	case score > 0:
		level = "low"
		label = "Niedrig"
	}

	return TrainingConfidence{
		Score: score,
		Level: level,
		Label: label,
	}
}

func trainingLabelConfidence(count int, target int) TrainingConfidence {
	if target <= 0 {
		target = 10
	}

	if count <= 0 {
		return trainingConfidenceFromScore(0)
	}

	// Grobe Datenabdeckung: target erreicht = 100%.
	// sqrt macht kleine Mengen etwas weniger hart, aber 1 Treffer bleibt niedrig.
	score := math.Sqrt(float64(count) / float64(target*2))

	return trainingConfidenceFromScore(score)
}

func trainingSaturationScore(value int, target int) float64 {
	if value <= 0 || target <= 0 {
		return 0
	}

	// Sanfter Anstieg, aber nie über 1.
	return clamp01(math.Sqrt(float64(value) / float64(target)))
}

func trainingAverageLabelConfidence(labels TrainingStatsLabels) float64 {
	values := []float64{}

	appendScores := func(items []TrainingLabelStat) {
		for _, item := range items {
			values = append(values, clamp01(item.Confidence.Score))
		}
	}

	appendScores(labels.People)
	appendScores(labels.SexPositions)
	appendScores(labels.BodyParts)
	appendScores(labels.Objects)
	appendScores(labels.Clothing)

	if len(values) == 0 {
		return 0
	}

	sum := 0.0
	for _, value := range values {
		sum += value
	}

	return clamp01(sum / float64(len(values)))
}

func trainingOverallConfidence(
	feedbackCount int,
	boxCount int,
	acceptedCount int,
	correctedCount int,
	labels TrainingStatsLabels,
) TrainingConfidence {
	if feedbackCount <= 0 {
		return trainingConfidenceFromScore(0)
	}

	// Datenmenge: 300 Feedbacks sind grob "voll", darunter anteilig.
	feedbackScore := trainingSaturationScore(feedbackCount, 300)

	// Detector-Daten: 1000 Boxen sind grob "voll", darunter anteilig.
	boxScore := trainingSaturationScore(boxCount, 1000)

	// Label-Abdeckung aus den einzelnen Label-Confidence-Werten.
	labelScore := trainingAverageLabelConfidence(labels)

	// Modell-/Prediction-Zustimmung:
	// Viele "Passt so"-Antworten bedeuten, dass die Vorhersagen brauchbar sind.
	// Bei 4/229 ist dieser Teil bewusst sehr niedrig.
	agreementScore := 0.0
	decisionCount := acceptedCount + correctedCount
	if decisionCount > 0 {
		agreementScore = clamp01(float64(acceptedCount) / float64(decisionCount))
	}

	// Korrekturquote als zusätzlicher Dämpfer.
	// 98% korrigiert soll die Gesamt-Confidence sichtbar drücken,
	// aber nicht alle gesammelten Daten entwerten.
	correctionRate := 0.0
	if decisionCount > 0 {
		correctionRate = clamp01(float64(correctedCount) / float64(decisionCount))
	}

	correctionPenalty := 1.0 - math.Min(0.45, correctionRate*0.45)

	// Gesamt:
	// - Datenmenge zählt viel
	// - Boxen und Label-Abdeckung zählen mittel
	// - echte Modell-Zustimmung zählt bewusst mit rein
	score :=
		feedbackScore*0.30 +
			boxScore*0.25 +
			labelScore*0.25 +
			agreementScore*0.20

	score *= correctionPenalty

	return trainingConfidenceFromScore(score)
}

func stringSet(values []string) map[string]bool {
	out := map[string]bool{}

	for _, value := range values {
		clean := strings.TrimSpace(value)
		if clean == "" {
			continue
		}

		out[clean] = true
	}

	return out
}

func trainingRecognitionEnabled() bool {
	return getSettings().TrainingRecognitionEnabled
}

func trainingDetectorEpochs() int {
	epochs := getSettings().TrainingDetectorEpochs

	if epochs < 1 {
		return 60
	}

	if epochs > 300 {
		return 300
	}

	return epochs
}

func trainingDeleteAllHandler(w http.ResponseWriter, r *http.Request) {
	if r.Method != http.MethodDelete && r.Method != http.MethodPost {
		trainingWriteError(w, http.StatusMethodNotAllowed, "method not allowed")
		return
	}

	root, err := trainingRootDir()
	if err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	if err := os.RemoveAll(root); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, fmt.Sprintf("Trainingsdaten konnten nicht gelöscht werden: %v", err))
		return
	}

	// Basisordner direkt wieder anlegen, damit die nächsten API-Aufrufe sauber funktionieren.
	if _, err := trainingRootDir(); err != nil {
		trainingWriteError(w, http.StatusInternalServerError, err.Error())
		return
	}

	trainingSetJobStatus(func(s *TrainingJobStatus) {
		*s = TrainingJobStatus{}
	})

	trainingWriteJSON(w, http.StatusOK, map[string]any{
		"ok":      true,
		"message": "Alle Trainingsdaten wurden gelöscht.",
	})
}

func trainingCountDetectorSamples(imagesDir string, labelsDir string) int {
	imageExts := map[string]bool{
		".jpg":  true,
		".jpeg": true,
		".png":  true,
		".webp": true,
	}

	entries, err := os.ReadDir(imagesDir)
	if err != nil {
		return 0
	}

	count := 0

	for _, e := range entries {
		if e.IsDir() {
			continue
		}

		ext := strings.ToLower(filepath.Ext(e.Name()))
		if !imageExts[ext] {
			continue
		}

		id := strings.TrimSuffix(e.Name(), filepath.Ext(e.Name()))
		labelPath := filepath.Join(labelsDir, id+".txt")

		if fileExists(labelPath) {
			count++
		}
	}

	return count
}

func trainingCountPositiveDetectorSamples(imagesDir string, labelsDir string) int {
	imageExts := map[string]bool{
		".jpg":  true,
		".jpeg": true,
		".png":  true,
		".webp": true,
	}

	entries, err := os.ReadDir(imagesDir)
	if err != nil {
		return 0
	}

	count := 0

	for _, e := range entries {
		if e.IsDir() {
			continue
		}

		ext := strings.ToLower(filepath.Ext(e.Name()))
		if !imageExts[ext] {
			continue
		}

		id := strings.TrimSuffix(e.Name(), filepath.Ext(e.Name()))
		labelPath := filepath.Join(labelsDir, id+".txt")

		if fileExistsNonEmpty(labelPath) {
			count++
		}
	}

	return count
}

func trainingWriteDetectorDatasetYAML(root string) error {
	datasetDir := filepath.Join(root, "detector", "dataset")

	absDatasetDir, err := filepath.Abs(datasetDir)
	if err != nil {
		return err
	}

	namesYAML, err := trainingDetectorDatasetNamesYAML()
	if err != nil {
		return err
	}

	path := filepath.Join(datasetDir, "dataset.yaml")
	yoloPath := filepath.ToSlash(absDatasetDir)

	body := fmt.Sprintf(`path: %s
train: images/train
val: images/val

names:
%s`, yoloPath, namesYAML)

	return os.WriteFile(path, []byte(body), 0644)
}

func trainingWritePoseDatasetYAML(root string) error {
	datasetDir := filepath.Join(root, "pose", "dataset")

	absDatasetDir, err := filepath.Abs(datasetDir)
	if err != nil {
		return err
	}

	namesYAML, err := trainingPoseDatasetNamesYAML()
	if err != nil {
		return err
	}

	path := filepath.Join(datasetDir, "dataset.yaml")
	yoloPath := filepath.ToSlash(absDatasetDir)

	body := fmt.Sprintf(`path: %s
train: images/train
val: images/val

kpt_shape: [17, 3]
flip_idx: [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15]

names:
%s`, yoloPath, namesYAML)

	return os.WriteFile(path, []byte(body), 0644)
}

func trainingEnsureDetectorDirs(root string) error {
	dirs := []string{
		filepath.Join(root, "detector"),
		filepath.Join(root, "detector", "dataset"),
		filepath.Join(root, "detector", "dataset", "images"),
		filepath.Join(root, "detector", "dataset", "images", "train"),
		filepath.Join(root, "detector", "dataset", "images", "val"),
		filepath.Join(root, "detector", "dataset", "labels"),
		filepath.Join(root, "detector", "dataset", "labels", "train"),
		filepath.Join(root, "detector", "dataset", "labels", "val"),
		filepath.Join(root, "detector", "runs"),
	}

	for _, dir := range dirs {
		if err := os.MkdirAll(dir, 0755); err != nil {
			return err
		}
	}

	if err := trainingWriteDetectorDatasetYAML(root); err != nil {
		return err
	}

	return nil
}

func trainingEnsurePoseDirs(root string) error {
	dirs := []string{
		filepath.Join(root, "pose"),
		filepath.Join(root, "pose", "dataset"),
		filepath.Join(root, "pose", "dataset", "images"),
		filepath.Join(root, "pose", "dataset", "images", "train"),
		filepath.Join(root, "pose", "dataset", "images", "val"),
		filepath.Join(root, "pose", "dataset", "labels"),
		filepath.Join(root, "pose", "dataset", "labels", "train"),
		filepath.Join(root, "pose", "dataset", "labels", "val"),
		filepath.Join(root, "pose", "runs"),
	}

	for _, dir := range dirs {
		if err := os.MkdirAll(dir, 0755); err != nil {
			return err
		}
	}

	if err := trainingWritePoseDatasetYAML(root); err != nil {
		return err
	}

	return nil
}

func trainingCreateNextSample() (*TrainingSample, error) {
	settings := getSettings()

	doneDir := strings.TrimSpace(settings.DoneDir)
	if doneDir == "" {
		return nil, errors.New("doneDir ist leer")
	}

	videoPath, err := trainingPickRandomVideo(doneDir)
	if err != nil {
		return nil, err
	}

	duration := trainingProbeDurationSeconds(videoPath)
	second := trainingRandomSecond(duration)

	root, err := trainingRootDir()
	if err != nil {
		return nil, err
	}

	if err := trainingEnsureDetectorDirs(root); err != nil {
		return nil, err
	}

	if err := os.MkdirAll(filepath.Join(root, "frames"), 0755); err != nil {
		return nil, err
	}
	if err := os.MkdirAll(filepath.Join(root, "samples"), 0755); err != nil {
		return nil, err
	}

	id := trainingMakeSampleID(videoPath, second)
	framePath := filepath.Join(root, "frames", id+".jpg")

	if err := trainingExtractFrame(videoPath, framePath, second); err != nil {
		// Fallback auf Sekunde 0, falls random second nicht funktioniert.
		second = 0
		id = trainingMakeSampleID(videoPath, second)
		framePath = filepath.Join(root, "frames", id+".jpg")

		if err2 := trainingExtractFrame(videoPath, framePath, second); err2 != nil {
			return nil, appErrorf("frame extraction failed: %v / fallback: %w", err, err2)
		}
	}

	prediction := trainingPredictFrame(framePath)

	var sourceSizeBytes int64
	if st, err := os.Stat(videoPath); err == nil && st != nil && !st.IsDir() {
		sourceSizeBytes = st.Size()
	}

	sample := &TrainingSample{
		SampleID:        id,
		FrameURL:        "/api/training/frame?id=" + id,
		SourceFile:      filepath.Base(videoPath),
		SourcePath:      videoPath,
		SourceSizeBytes: sourceSizeBytes,
		Second:          second,
		CreatedAt:       time.Now().UTC().Format(time.RFC3339),
		Prediction:      prediction,
	}

	if err := trainingWriteSample(root, sample); err != nil {
		return nil, err
	}

	return sample, nil
}

func trainingCreateUncertainNextSampleWithProgress(startedAtMs int64, requestID string) (*TrainingSample, error) {
	root, err := trainingRootDir()
	if err != nil {
		return nil, err
	}

	candidates, err := trainingReadValidUncertainCandidates(root)
	if err != nil {
		return nil, err
	}

	// Ziel:
	// Vor jeder Auswahl sollen wieder 5 Kandidaten verglichen werden.
	// Wenn noch 4 aus der Queue übrig sind, wird genau 1 neues Bild extrahiert.
	// Wenn keine Queue existiert, wird ein kompletter 5er-Batch aufgebaut.
	missing := trainingUncertainCandidateCount - len(candidates)
	if missing < 1 {
		// Trotzdem mindestens 1 neues Bild nachziehen,
		// damit die Auswahl nach jedem Speichern frisch bleibt.
		missing = 1
	}

	candidateWindowTotal := trainingUncertainCandidateCount

	const stepsPerCandidate = 4
	totalSteps := missing*stepsPerCandidate + 1

	errs := []string{}

	for i := 0; i < missing; i++ {
		attempt := i + 1
		stepStart := i*stepsPerCandidate + 1

		prefix := fmt.Sprintf("Kandidat %d/%d: ", attempt, candidateWindowTotal)

		candidate, err := trainingCreateUncertainCandidateWithProgress(
			root,
			startedAtMs,
			requestID,
			stepStart,
			totalSteps,
			prefix,
		)

		if err != nil {
			errs = append(errs, err.Error())

			trainingPublishAnalysisStep(
				requestID,
				startedAtMs,
				stepStart+stepsPerCandidate-1,
				totalSteps,
				"",
				fmt.Sprintf("Kandidat %d/%d fehlgeschlagen…", attempt, candidateWindowTotal),
			)

			continue
		}

		candidates = append(candidates, *candidate)

		trainingPublishAnalysisStep(
			requestID,
			startedAtMs,
			stepStart+stepsPerCandidate-1,
			totalSteps,
			candidate.sample.SourceFile,
			fmt.Sprintf(
				"Kandidat %d/%d bewertet · Unsicherheit %.0f%%",
				attempt,
				candidateWindowTotal,
				candidate.score*100,
			),
		)
	}

	if len(candidates) == 0 {
		if len(errs) > 0 {
			return nil, errors.New(strings.Join(errs, "; "))
		}

		return nil, errors.New("keine unsicheren Trainingskandidaten gefunden")
	}

	sort.Slice(candidates, func(i, j int) bool {
		if candidates[i].score == candidates[j].score {
			return candidates[i].sample.CreatedAt < candidates[j].sample.CreatedAt
		}

		return candidates[i].score > candidates[j].score
	})

	best := candidates[0]
	remaining := candidates[1:]

	if len(remaining) > trainingUncertainCandidateCount-1 {
		remaining = remaining[:trainingUncertainCandidateCount-1]
	}

	if err := trainingWriteUncertainCandidateQueue(root, remaining); err != nil {
		return nil, err
	}

	trainingPublishAnalysisStep(
		requestID,
		startedAtMs,
		totalSteps,
		totalSteps,
		best.sample.SourceFile,
		fmt.Sprintf(
			"Unsicherster Kandidat gewählt · Score %.0f%% · Fenster %d/%d",
			best.score*100,
			len(remaining)+1,
			trainingUncertainCandidateCount,
		),
	)

	best.sample.UncertaintyScore = best.score

	return best.sample, nil
}

func reloadAIServerModelAfterTraining() {
	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
	defer cancel()

	req, err := http.NewRequestWithContext(ctx, http.MethodPost, aiServerURL()+"/reload", nil)
	if err != nil {
		appLogln("⚠️ AI Server Reload Request konnte nicht gebaut werden:", err)
		return
	}

	addAIServerAuth(req)

	client := &http.Client{
		Timeout: 30 * time.Second,
	}

	resp, err := client.Do(req)
	if err != nil {
		appLogln("⚠️ AI Server Reload fehlgeschlagen:", err)
		return
	}
	defer resp.Body.Close()

	body, _ := io.ReadAll(io.LimitReader(resp.Body, 4096))

	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
		appLogln("⚠️ AI Server Reload HTTP", resp.StatusCode, strings.TrimSpace(string(body)))
		return
	}

	appLogln("✅ AI Server Modell nach Training neu geladen:", strings.TrimSpace(string(body)))
}

func trainingDeleteSampleFiles(root string, sampleID string) {
	sampleID = strings.TrimSpace(sampleID)
	if sampleID == "" || strings.Contains(sampleID, "/") || strings.Contains(sampleID, "\\") {
		return
	}

	_ = os.Remove(filepath.Join(root, "samples", sampleID+".json"))
	_ = os.Remove(filepath.Join(root, "frames", sampleID+".jpg"))
}

func trainingPredictionUncertaintyScore(pred TrainingPrediction) float64 {
	if !pred.ModelAvailable {
		return 0.10
	}

	scores := []float64{}

	addScore := func(score float64) {
		if math.IsNaN(score) || math.IsInf(score, 0) {
			return
		}

		if score <= 0 {
			return
		}

		scores = append(scores, clamp01(score))
	}

	if !isNoSexPositionLabel(pred.SexPosition) {
		addScore(pred.SexPositionScore)
	}

	for _, box := range pred.Boxes {
		addScore(box.Score)
	}

	if len(pred.Boxes) == 0 {
		for _, item := range pred.BodyPartsPresent {
			addScore(item.Score)
		}

		for _, item := range pred.ObjectsPresent {
			addScore(item.Score)
		}

		for _, item := range pred.ClothingPresent {
			addScore(item.Score)
		}
	}

	if len(scores) == 0 {
		// Modell ist verfügbar, erkennt aber nichts.
		// Das kann ein nützliches Hard-Negative oder ein False-Negative sein.
		return 0.35
	}

	sum := 0.0

	for _, score := range scores {
		// Höchste Unsicherheit ungefähr im mittleren Bereich.
		// 0.55 ist absichtlich etwas niedriger als 0.75,
		// damit Low/Mid-Confidence-Fälle bevorzugt werden.
		distance := math.Abs(score - 0.55)
		uncertainty := 1.0 - distance/0.55
		sum += clamp01(uncertainty)
	}

	avg := sum / float64(len(scores))

	// Viele Boxen mit mittlerer Confidence sind besonders wertvoll.
	boxBonus := math.Min(0.12, float64(len(pred.Boxes))*0.025)

	// Sehr niedrige Confidence soll ebenfalls auftauchen,
	// aber nicht alle Ergebnisse dominieren.
	lowConfidenceBonus := 0.0
	for _, score := range scores {
		if score >= 0.25 && score <= 0.75 {
			lowConfidenceBonus = 0.08
			break
		}
	}

	return clamp01(avg + boxBonus + lowConfidenceBonus)
}

func trainingCreateNextSampleWithProgress(startedAtMs int64, requestID string) (*TrainingSample, error) {
	return trainingCreateNextSampleWithProgressRange(
		startedAtMs,
		requestID,
		1,
		4,
		"",
	)
}

func trainingCreateNextSampleWithProgressRange(
	startedAtMs int64,
	requestID string,
	stepStart int,
	stepTotal int,
	prefix string,
) (*TrainingSample, error) {
	publishStep := func(localStep int, sourceFile string, message string) {
		trainingPublishAnalysisStep(
			requestID,
			startedAtMs,
			stepStart+localStep-1,
			stepTotal,
			sourceFile,
			prefix+message,
		)
	}

	publishStep(1, "", "Video wird ausgewählt…")

	settings := getSettings()

	doneDir := strings.TrimSpace(settings.DoneDir)
	if doneDir == "" {
		return nil, errors.New("doneDir ist leer")
	}

	videoPath, err := trainingPickRandomVideo(doneDir)
	if err != nil {
		return nil, err
	}

	sourceFile := filepath.Base(videoPath)
	previewURL := ""

	publishStep = func(localStep int, sourceFile string, message string) {
		trainingPublishAnalysisStepWithPreview(
			requestID,
			startedAtMs,
			stepStart+localStep-1,
			stepTotal,
			sourceFile,
			previewURL,
			prefix+message,
		)
	}

	publishStep(2, sourceFile, "Bild wird extrahiert…")

	duration := trainingProbeDurationSeconds(videoPath)
	second := trainingRandomSecond(duration)

	root, err := trainingRootDir()
	if err != nil {
		return nil, err
	}

	if err := trainingEnsureDetectorDirs(root); err != nil {
		return nil, err
	}

	if err := os.MkdirAll(filepath.Join(root, "frames"), 0755); err != nil {
		return nil, err
	}

	if err := os.MkdirAll(filepath.Join(root, "samples"), 0755); err != nil {
		return nil, err
	}

	id := trainingMakeSampleID(videoPath, second)
	framePath := filepath.Join(root, "frames", id+".jpg")

	if err := trainingExtractFrame(videoPath, framePath, second); err != nil {
		publishStep(2, sourceFile, "Bild wird erneut bei Sekunde 0 extrahiert…")

		second = 0
		id = trainingMakeSampleID(videoPath, second)
		framePath = filepath.Join(root, "frames", id+".jpg")

		if err2 := trainingExtractFrame(videoPath, framePath, second); err2 != nil {
			return nil, appErrorf("frame extraction failed: %v / fallback: %w", err, err2)
		}
	}

	previewURL = "/api/training/frame?id=" + url.QueryEscape(id)

	publishStep(3, sourceFile, "Bild wird analysiert…")

	prediction := trainingPredictFrame(framePath)

	var sourceSizeBytes int64
	if st, err := os.Stat(videoPath); err == nil && st != nil && !st.IsDir() {
		sourceSizeBytes = st.Size()
	}

	sample := &TrainingSample{
		SampleID:        id,
		FrameURL:        "/api/training/frame?id=" + id,
		SourceFile:      sourceFile,
		SourcePath:      videoPath,
		SourceSizeBytes: sourceSizeBytes,
		Second:          second,
		CreatedAt:       time.Now().UTC().Format(time.RFC3339),
		Prediction:      prediction,
	}

	publishStep(4, sourceFile, "Analyse-Ergebnis wird gespeichert…")

	if err := trainingWriteSample(root, sample); err != nil {
		return nil, err
	}

	return sample, nil
}

func trainingPickRandomVideo(doneDir string) (string, error) {
	extOK := map[string]bool{
		".mp4": true,
	}

	doneDir = strings.TrimSpace(doneDir)
	if doneDir == "" {
		return "", errors.New("doneDir ist leer")
	}

	doneAbs, err := filepath.Abs(doneDir)
	if err != nil {
		return "", err
	}

	var files []string

	err = filepath.WalkDir(doneAbs, func(path string, d os.DirEntry, walkErr error) error {
		if walkErr != nil {
			return nil
		}

		rel, err := filepath.Rel(doneAbs, path)
		if err != nil {
			return nil
		}

		rel = filepath.Clean(rel)

		// Root "done" selbst erlauben.
		if rel == "." {
			return nil
		}

		parts := strings.Split(rel, string(os.PathSeparator))
		top := strings.ToLower(strings.TrimSpace(parts[0]))
		name := strings.ToLower(strings.TrimSpace(d.Name()))

		if d.IsDir() {
			// Nur diese Bereiche fürs Trainingsbild:
			// - done/
			// - done/keep/...
			//
			// Alles andere unter done wird ignoriert, z.B.:
			// .postwork_tmp, .trash, generated, training, sonstige Temp-Ordner.
			if top != "keep" {
				return filepath.SkipDir
			}

			// Innerhalb von keep trotzdem versteckte/temporäre Ordner überspringen.
			if name == ".trash" ||
				name == ".postwork_tmp" ||
				name == "training" ||
				name == "generated" ||
				strings.HasPrefix(name, ".") {
				return filepath.SkipDir
			}

			return nil
		}

		// Dateien nur direkt in done/ oder unter done/keep/... erlauben.
		if len(parts) > 1 && top != "keep" {
			return nil
		}

		// Keine versteckten/temp-Dateien verwenden.
		if strings.HasPrefix(name, ".") ||
			strings.Contains(name, ".tmp.") ||
			strings.Contains(name, ".part") {
			return nil
		}

		ext := strings.ToLower(filepath.Ext(path))
		if extOK[ext] {
			files = append(files, path)
		}

		return nil
	})

	if err != nil {
		return "", err
	}

	if len(files) == 0 {
		return "", errors.New("keine Videos in done oder done/keep gefunden")
	}

	return files[rand.Intn(len(files))], nil
}

func trainingExtractFrame(videoPath string, framePath string, second float64) error {
	settings := getSettings()

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

	_ = os.Remove(framePath)

	ss := strconv.FormatFloat(math.Max(0, second), 'f', 3, 64)

	cmd := exec.Command(
		ffmpeg,
		"-hide_banner",
		"-loglevel", "error",
		"-ss", ss,
		"-i", videoPath,
		"-frames:v", "1",
		"-q:v", "2",
		"-y",
		framePath,
	)

	cmd.SysProcAttr = &syscall.SysProcAttr{
		HideWindow:    true,
		CreationFlags: 0x08000000,
	}

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

	if st, err := os.Stat(framePath); err != nil || st.Size() == 0 {
		return errors.New("ffmpeg erzeugte kein Frame")
	}

	return nil
}

func trainingPredictFrame(framePath string) TrainingPrediction {
	settings := getSettings()
	if !settings.TrainingRecognitionEnabled {
		return trainingEmptyPrediction("recognition_disabled")
	}

	root, err := trainingRootDir()
	if err != nil {
		appLogln("⚠️ training predict root error:", err)
		return trainingEmptyPrediction("root_error")
	}

	det := trainingPredictDetector(root, framePath)
	pred := trainingPredictionFromDetector(det)

	pose := trainingPredictPose(root, framePath)
	pred = trainingApplyPoseToPrediction(pred, pose)

	return pred
}

func trainingPredictFrameDetectorOnly(framePath string) TrainingPrediction {
	settings := getSettings()
	if !settings.TrainingRecognitionEnabled {
		return trainingEmptyPrediction("recognition_disabled")
	}

	root, err := trainingRootDir()
	if err != nil {
		appLogln("⚠️ training predict root error:", err)
		return trainingEmptyPrediction("root_error")
	}

	det := trainingPredictDetector(root, framePath)
	return trainingPredictionFromDetector(det)
}

func trainingPredictionFromDetector(det TrainingDetectorPrediction) TrainingPrediction {
	rawBoxes := det.Boxes
	if rawBoxes == nil {
		rawBoxes = []TrainingBox{}
	}

	pred := TrainingPrediction{
		ModelAvailable:   det.Available,
		Source:           det.Source,
		SexPosition:      trainingNoSexPositionLabel,
		SexPositionScore: 0,
		PeoplePresent:    []TrainingScoredLabel{},
		BodyPartsPresent: []TrainingScoredLabel{},
		ObjectsPresent:   []TrainingScoredLabel{},
		ClothingPresent:  []TrainingScoredLabel{},
		Boxes:            []TrainingBox{},
	}

	if pred.Source == "" {
		if det.Available {
			pred.Source = "yolo26_detector"
		} else {
			pred.Source = "detector_missing"
		}
	}

	grouped, err := trainingGroupedLabels()
	if err != nil {
		appLogln("⚠️ detector label grouping failed:", err)
		return pred
	}

	peopleSet := map[string]bool{}
	for _, label := range grouped.People {
		clean := strings.TrimSpace(label)
		if clean != "" {
			peopleSet[clean] = true
		}
	}

	detectionSet := map[string]bool{}

	for _, label := range grouped.BodyParts {
		clean := strings.TrimSpace(label)
		if clean != "" {
			detectionSet[clean] = true
		}
	}

	for _, label := range grouped.Objects {
		clean := strings.TrimSpace(label)
		if clean != "" {
			detectionSet[clean] = true
		}
	}

	for _, label := range grouped.Clothing {
		clean := strings.TrimSpace(label)
		if clean != "" {
			detectionSet[clean] = true
		}
	}

	visibleBoxes := []TrainingBox{}

	for _, box := range rawBoxes {
		if box.Score > 0 && box.Score < 0.25 {
			continue
		}

		label := strings.TrimSpace(box.Label)
		if label == "" {
			continue
		}

		box.Label = label

		if peopleSet[label] {
			visibleBoxes = append(visibleBoxes, box)
			continue
		}

		if detectionSet[label] {
			visibleBoxes = append(visibleBoxes, box)
		}
	}

	pred.Boxes = visibleBoxes

	pred.PeoplePresent = trainingScoredLabelsFromDetectorBoxes(visibleBoxes, grouped.People)
	pred.BodyPartsPresent = trainingScoredLabelsFromDetectorBoxes(visibleBoxes, grouped.BodyParts)
	pred.ObjectsPresent = trainingScoredLabelsFromDetectorBoxes(visibleBoxes, grouped.Objects)
	pred.ClothingPresent = trainingScoredLabelsFromDetectorBoxes(visibleBoxes, grouped.Clothing)

	return pred
}

func trainingPredictDetector(root string, framePath string) TrainingDetectorPrediction {
	python := trainingPythonExe()
	script := trainingScriptPath("predict_detector_model.py")

	model := trainingResolveDetectorModel(root)
	if !model.EffectiveExists {
		return TrainingDetectorPrediction{
			Available: false,
			Source:    model.Source,
			Boxes:     []TrainingBox{},
		}
	}

	confValues := []string{"0.30"}

	best := TrainingDetectorPrediction{
		Available: true,
		Source:    "yolo26_detector",
		Boxes:     []TrainingBox{},
	}

	for _, conf := range confValues {
		cmd := exec.Command(
			python,
			script,
			"--root", root,
			"--image", framePath,
			"--model", model.EffectivePath,
			"--conf", conf,
			"--imgsz", "640",
		)

		cmd.SysProcAttr = &syscall.SysProcAttr{
			HideWindow:    true,
			CreationFlags: 0x08000000,
		}

		var stdout strings.Builder
		var stderr strings.Builder

		cmd.Stdout = &stdout
		cmd.Stderr = &stderr

		err := cmd.Run()

		outText := strings.TrimSpace(stdout.String())
		errText := strings.TrimSpace(stderr.String())

		if errText != "" {
			appLogln("🔎 detector stderr:", errText)
		}

		if err != nil {
			appLogln("⚠️ detector predict failed")
			appLogln("  conf:", conf)
			appLogln("  error:", err)
			appLogln("  stdout:", outText)
			appLogln("  stderr:", errText)
			continue
		}

		if outText == "" {
			appLogln("⚠️ detector predict empty stdout")
			appLogln("  conf:", conf)
			appLogln("  stderr:", errText)
			continue
		}

		var det TrainingDetectorPrediction
		if err := json.Unmarshal([]byte(outText), &det); err != nil {
			appLogln("⚠️ detector predict json failed:", err)
			appLogln("  conf:", conf)
			appLogln("  stdout:", outText)
			appLogln("  stderr:", errText)
			continue
		}

		if det.Boxes == nil {
			det.Boxes = []TrainingBox{}
		}

		det.Source = model.Source

		best = det

		if len(det.Boxes) > 0 {
			return det
		}
	}

	if best.Boxes == nil {
		best.Boxes = []TrainingBox{}
	}

	return best
}

func trainingScoredLabelsFromDetectorBoxes(
	boxes []TrainingBox,
	group []string,
) []TrainingScoredLabel {
	groupSet := map[string]bool{}

	for _, label := range group {
		clean := strings.TrimSpace(label)
		if clean != "" {
			groupSet[clean] = true
		}
	}

	best := map[string]float64{}

	for _, box := range boxes {
		label := strings.TrimSpace(box.Label)
		if label == "" || !groupSet[label] {
			continue
		}

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

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

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

	for label, score := range best {
		out = append(out, TrainingScoredLabel{
			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 trainingApplyDetectorToPrediction(pred TrainingPrediction, det TrainingDetectorPrediction) TrainingPrediction {
	boxes := det.Boxes
	if boxes == nil {
		boxes = []TrainingBox{}
	}

	grouped, err := trainingGroupedLabels()
	if err != nil {
		appLogln("⚠️ detector label grouping failed:", err)

		pred.Boxes = boxes
		pred.BodyPartsPresent = []TrainingScoredLabel{}
		pred.ObjectsPresent = []TrainingScoredLabel{}
		pred.ClothingPresent = []TrainingScoredLabel{}
		return pred
	}

	// Wichtig:
	// Ab jetzt kommen diese drei Bereiche ausschließlich vom Object Detector.
	// Kein Scene-KNN-Fallback, damit keine Labels ohne Box angezeigt werden.
	pred.BodyPartsPresent = trainingScoredLabelsFromDetectorBoxes(boxes, grouped.BodyParts)
	pred.ObjectsPresent = trainingScoredLabelsFromDetectorBoxes(boxes, grouped.Objects)
	pred.ClothingPresent = trainingScoredLabelsFromDetectorBoxes(boxes, grouped.Clothing)
	pred.Boxes = boxes

	if det.Available {
		if pred.Source == "" {
			pred.Source = "yolo26_detector"
		} else {
			pred.Source = pred.Source + "+yolo_detector"
		}
		pred.ModelAvailable = true
	}

	return pred
}

func trainingPredictPose(root string, framePath string) TrainingPosePrediction {
	python := trainingPythonExe()
	script := trainingScriptPath("predict_pose_model.py")

	model := trainingResolvePoseModel(root)
	if !model.EffectiveExists {
		return TrainingPosePrediction{
			Available: false,
			Source:    model.Source,
			Persons:   []TrainingPosePerson{},
		}
	}

	cmd := exec.Command(
		python,
		script,
		"--root", root,
		"--image", framePath,
		"--model", model.EffectivePath,
		"--conf", "0.30",
		"--imgsz", "640",
	)

	cmd.SysProcAttr = &syscall.SysProcAttr{
		HideWindow:    true,
		CreationFlags: 0x08000000,
	}

	var stdout strings.Builder
	var stderr strings.Builder

	cmd.Stdout = &stdout
	cmd.Stderr = &stderr

	err := cmd.Run()

	outText := strings.TrimSpace(stdout.String())
	errText := strings.TrimSpace(stderr.String())

	if errText != "" {
		appLogln("🔎 pose stderr:", errText)
	}

	if err != nil {
		appLogln("⚠️ pose predict failed:", err)
		appLogln("  stdout:", outText)
		return TrainingPosePrediction{
			Available: false,
			Source:    "pose_error",
			Persons:   []TrainingPosePerson{},
		}
	}

	if outText == "" {
		return TrainingPosePrediction{
			Available: false,
			Source:    "pose_empty",
			Persons:   []TrainingPosePerson{},
		}
	}

	var pose TrainingPosePrediction
	if err := json.Unmarshal([]byte(outText), &pose); err != nil {
		appLogln("⚠️ pose predict json failed:", err)
		return TrainingPosePrediction{
			Available: false,
			Source:    "pose_json_error",
			Persons:   []TrainingPosePerson{},
		}
	}

	if pose.Persons == nil {
		pose.Persons = []TrainingPosePerson{}
	}

	pose = trainingAnnotatePosePrediction(pose)
	pose.Source = model.Source

	return pose
}

func trainingCombinePositionScore(scores map[string]float64, positionSet map[string]bool, label string, score float64) {
	label = normalizeSexPositionLabel(label)
	if isNoSexPositionLabel(label) || !positionSet[label] || score <= 0 {
		return
	}

	score = clamp01(score)
	current := clamp01(scores[label])
	scores[label] = clamp01(1 - (1-current)*(1-score))
}

func trainingAppendPredictionSource(pred *TrainingPrediction, source string) {
	source = strings.TrimSpace(source)
	if source == "" {
		return
	}

	current := strings.TrimSpace(pred.Source)
	if current == "" {
		pred.Source = source
		return
	}

	for _, part := range strings.Split(current, "+") {
		if strings.TrimSpace(part) == source {
			return
		}
	}

	pred.Source = current + "+" + source
}

func trainingFuseHybridPositionScores(
	poseScores map[string]float64,
	contextScores map[string]float64,
) (string, float64, bool, bool) {
	labels := map[string]bool{}

	for label := range poseScores {
		if !isNoSexPositionLabel(label) {
			labels[label] = true
		}
	}

	for label := range contextScores {
		if !isNoSexPositionLabel(label) {
			labels[label] = true
		}
	}

	bestPosition := ""
	bestPositionScore := 0.0
	bestHasPose := false
	bestHasContext := false
	bestPosePosition := ""
	bestPoseScore := 0.0
	bestPoseHasContext := false

	for label := range labels {
		poseScore := clamp01(poseScores[label])
		contextScore := clamp01(contextScores[label])

		score := 0.0
		hasPose := poseScore > 0
		hasContext := contextScore > 0

		if hasPose {
			score = poseScore

			// Kontext boostet die Pose, bleibt aber bewusst Nebeninformation.
			if hasContext {
				boost := clamp01(contextScore * trainingPositionContextBoostWeight)
				score = clamp01(1 - (1-score)*(1-boost))
			}
		} else if contextScore >= trainingPositionContextMinScore {
			// Reiner Box-/Szenen-Kontext darf eine unsichere Prediction liefern,
			// aber keine hohe Sicherheit vortaeuschen.
			score = math.Min(trainingPositionContextMaxScore, contextScore)
		}

		if score > bestPositionScore {
			bestPosition = label
			bestPositionScore = score
			bestHasPose = hasPose
			bestHasContext = hasContext
		}

		if hasPose && score > bestPoseScore {
			bestPosePosition = label
			bestPoseScore = score
			bestPoseHasContext = hasContext
		}
	}

	if bestPosePosition != "" &&
		!bestHasPose &&
		bestPositionScore <= bestPoseScore+trainingPositionContextOverrideMargin {
		bestPosition = bestPosePosition
		bestPositionScore = bestPoseScore
		bestHasPose = true
		bestHasContext = bestPoseHasContext
	}

	return bestPosition, clamp01(bestPositionScore), bestHasPose, bestHasContext
}

func trainingIsFinite01(v float64) bool {
	return !math.IsNaN(v) && !math.IsInf(v, 0) && v >= 0 && v <= 1
}

func trainingNormalizedBox(box TrainingBox) (TrainingBox, bool) {
	x := clamp01(box.X)
	y := clamp01(box.Y)
	w := clamp01(box.W)
	h := clamp01(box.H)

	if w <= 0 || h <= 0 {
		return TrainingBox{}, false
	}

	if x+w > 1 {
		w = 1 - x
	}
	if y+h > 1 {
		h = 1 - y
	}

	if w <= 0 || h <= 0 {
		return TrainingBox{}, false
	}

	box.X = x
	box.Y = y
	box.W = w
	box.H = h
	return box, true
}

func trainingBoxArea(box TrainingBox) float64 {
	box, ok := trainingNormalizedBox(box)
	if !ok {
		return 0
	}

	return box.W * box.H
}

func trainingBoxCenter(box TrainingBox) (float64, float64, bool) {
	box, ok := trainingNormalizedBox(box)
	if !ok {
		return 0, 0, false
	}

	return box.X + box.W/2, box.Y + box.H/2, true
}

func trainingBoxGap(a TrainingBox, b TrainingBox) float64 {
	a, okA := trainingNormalizedBox(a)
	b, okB := trainingNormalizedBox(b)
	if !okA || !okB {
		return 1
	}

	dx := math.Max(0, math.Max(a.X-(b.X+b.W), b.X-(a.X+a.W)))
	dy := math.Max(0, math.Max(a.Y-(b.Y+b.H), b.Y-(a.Y+a.H)))

	return math.Sqrt(dx*dx + dy*dy)
}

func trainingBoxOverlapRatio(a TrainingBox, b TrainingBox) float64 {
	a, okA := trainingNormalizedBox(a)
	b, okB := trainingNormalizedBox(b)
	if !okA || !okB {
		return 0
	}

	left := math.Max(a.X, b.X)
	top := math.Max(a.Y, b.Y)
	right := math.Min(a.X+a.W, b.X+b.W)
	bottom := math.Min(a.Y+a.H, b.Y+b.H)

	if right <= left || bottom <= top {
		return 0
	}

	intersection := (right - left) * (bottom - top)
	minArea := math.Min(trainingBoxArea(a), trainingBoxArea(b))
	if minArea <= 0 {
		return 0
	}

	return clamp01(intersection / minArea)
}

func trainingBoxHorizontalOverlapRatio(a TrainingBox, b TrainingBox) float64 {
	a, okA := trainingNormalizedBox(a)
	b, okB := trainingNormalizedBox(b)
	if !okA || !okB {
		return 0
	}

	left := math.Max(a.X, b.X)
	right := math.Min(a.X+a.W, b.X+b.W)
	if right <= left {
		return 0
	}

	minWidth := math.Min(a.W, b.W)
	if minWidth <= 0 {
		return 0
	}

	return clamp01((right - left) / minWidth)
}

func trainingBoxesByLabel(boxes []TrainingBox, labels ...string) []TrainingBox {
	wanted := map[string]bool{}
	for _, label := range labels {
		clean := strings.TrimSpace(label)
		if clean != "" {
			wanted[clean] = true
		}
	}

	out := []TrainingBox{}
	for _, box := range boxes {
		label := strings.TrimSpace(box.Label)
		if !wanted[label] {
			continue
		}
		if normalized, ok := trainingNormalizedBox(box); ok {
			out = append(out, normalized)
		}
	}

	return out
}

func trainingAnyBoxesNear(a []TrainingBox, b []TrainingBox, margin float64) bool {
	for _, left := range a {
		for _, right := range b {
			if trainingBoxGap(left, right) <= margin || trainingBoxOverlapRatio(left, right) > 0 {
				return true
			}
		}
	}

	return false
}

func trainingPointNearBox(x float64, y float64, box TrainingBox, margin float64) bool {
	if !trainingIsFinite01(x) || !trainingIsFinite01(y) {
		return false
	}

	box, ok := trainingNormalizedBox(box)
	if !ok {
		return false
	}

	return x >= box.X-margin &&
		x <= box.X+box.W+margin &&
		y >= box.Y-margin &&
		y <= box.Y+box.H+margin
}

func trainingAnyPoseKeypointNearBoxes(
	pose TrainingPosePrediction,
	keypointNames []string,
	boxes []TrainingBox,
	margin float64,
) bool {
	if len(boxes) == 0 {
		return false
	}

	nameSet := stringSet(keypointNames)

	for _, person := range pose.Persons {
		if !trainingPosePersonReliable(person) {
			continue
		}

		for _, point := range person.Keypoints {
			if !nameSet[strings.TrimSpace(point.Name)] {
				continue
			}
			if point.Conf < trainingPoseKeypointMinConfidence {
				continue
			}

			for _, box := range boxes {
				if trainingPointNearBox(point.X, point.Y, box, margin) {
					return true
				}
			}
		}
	}

	return false
}

func trainingPoseKeypointStats(person TrainingPosePerson) (int, float64) {
	if len(person.Keypoints) == 0 {
		return 0, 0
	}

	visible := 0
	totalConf := 0.0

	for _, point := range person.Keypoints {
		if point.Conf < trainingPoseKeypointMinConfidence ||
			!trainingIsFinite01(point.X) ||
			!trainingIsFinite01(point.Y) {
			continue
		}

		visible++
		totalConf += clamp01(point.Conf)
	}

	if visible == 0 {
		return 0, 0
	}

	coverage := clamp01(float64(visible) / float64(trainingPoseKeypointCount))
	avgConf := clamp01(totalConf / float64(visible))

	return visible, clamp01(coverage*0.45 + avgConf*0.55)
}

func trainingPoseKeypointQuality(person TrainingPosePerson) float64 {
	_, quality := trainingPoseKeypointStats(person)
	return quality
}

func trainingAnnotatePosePersonQuality(person TrainingPosePerson) TrainingPosePerson {
	visible, quality := trainingPoseKeypointStats(person)
	person.VisibleKeypoints = visible
	person.Quality = quality
	person.Reliable = person.Score >= trainingPoseReliableMinScore &&
		visible >= trainingPoseReliableMinKeypoints &&
		quality >= trainingPoseReliableMinQuality

	return person
}

func trainingPosePersonReliable(person TrainingPosePerson) bool {
	visible := person.VisibleKeypoints
	quality := person.Quality

	if visible == 0 && quality == 0 && len(person.Keypoints) > 0 {
		visible, quality = trainingPoseKeypointStats(person)
	}

	return person.Score >= trainingPoseReliableMinScore &&
		visible >= trainingPoseReliableMinKeypoints &&
		quality >= trainingPoseReliableMinQuality
}

func trainingAnnotatePosePrediction(pose TrainingPosePrediction) TrainingPosePrediction {
	for i := range pose.Persons {
		pose.Persons[i] = trainingAnnotatePosePersonQuality(pose.Persons[i])
	}

	if pose.PersonCount == 0 {
		pose.PersonCount = len(pose.Persons)
	}

	return pose
}

func trainingReliablePosePrediction(pose TrainingPosePrediction) TrainingPosePrediction {
	pose = trainingAnnotatePosePrediction(pose)

	reliable := make([]TrainingPosePerson, 0, len(pose.Persons))
	for _, person := range pose.Persons {
		if trainingPosePersonReliable(person) {
			reliable = append(reliable, person)
		}
	}

	pose.Persons = reliable
	pose.PersonCount = len(reliable)

	return pose
}

func trainingScenePersonBoxes(pred TrainingPrediction, pose TrainingPosePrediction) []TrainingBox {
	detectorBoxes := []TrainingBox{}
	poseBoxes := []TrainingBox{}

	for _, box := range pred.Boxes {
		if trainingIsPersonLikeLabel(box.Label) {
			if normalized, ok := trainingNormalizedBox(box); ok {
				detectorBoxes = append(detectorBoxes, normalized)
			}
		}
	}

	for _, person := range pose.Persons {
		if !trainingPosePersonReliable(person) {
			continue
		}

		if normalized, ok := trainingNormalizedBox(person.Box); ok {
			poseBoxes = append(poseBoxes, normalized)
		}
	}

	if len(detectorBoxes) >= len(poseBoxes) && len(detectorBoxes) > 0 {
		return detectorBoxes
	}

	return poseBoxes
}

func trainingIsPersonLikeLabel(label string) bool {
	label = strings.TrimSpace(label)
	return label == "person" || strings.HasPrefix(label, "person_")
}

type trainingPersonPairSignals struct {
	close      bool
	overlap    bool
	horizontal bool
	vertical   bool
	stacked    bool
}

func trainingScenePersonPairSignals(boxes []TrainingBox) trainingPersonPairSignals {
	signals := trainingPersonPairSignals{}

	for i := 0; i < len(boxes); i++ {
		a, okA := trainingNormalizedBox(boxes[i])
		if !okA {
			continue
		}

		for j := i + 1; j < len(boxes); j++ {
			b, okB := trainingNormalizedBox(boxes[j])
			if !okB {
				continue
			}

			gap := trainingBoxGap(a, b)
			overlap := trainingBoxOverlapRatio(a, b)
			close := gap <= 0.08 || overlap >= 0.08
			if close {
				signals.close = true
			}
			if overlap >= 0.18 {
				signals.overlap = true
			}

			if close && a.W > a.H*1.15 && b.W > b.H*1.15 {
				signals.horizontal = true
			}
			if close && a.H > a.W*1.25 && b.H > b.W*1.25 {
				signals.vertical = true
			}

			_, ay, okACenter := trainingBoxCenter(a)
			_, by, okBCenter := trainingBoxCenter(b)
			if okACenter && okBCenter &&
				gap <= 0.12 &&
				trainingBoxHorizontalOverlapRatio(a, b) >= 0.35 &&
				math.Abs(ay-by) >= 0.12 {
				signals.stacked = true
			}
		}
	}

	return signals
}

type trainingPosePoint struct {
	x float64
	y float64
}

type trainingPosePersonGeometry struct {
	box            TrainingBox
	hasBox         bool
	center         trainingPosePoint
	lying          bool
	upright        bool
	straddling     bool
	kneesBelowHips bool
	kneesWide      bool
	allFours       bool
	bentOrKneeling bool
}

func trainingPosePointByName(person TrainingPosePerson, name string) (trainingPosePoint, bool) {
	name = strings.TrimSpace(name)
	if name == "" {
		return trainingPosePoint{}, false
	}

	for _, point := range person.Keypoints {
		if strings.TrimSpace(point.Name) != name {
			continue
		}
		if point.Conf < trainingPoseKeypointMinConfidence ||
			!trainingIsFinite01(point.X) ||
			!trainingIsFinite01(point.Y) {
			continue
		}

		return trainingPosePoint{x: point.X, y: point.Y}, true
	}

	return trainingPosePoint{}, false
}

func trainingPoseMidpoint(person TrainingPosePerson, names ...string) (trainingPosePoint, bool) {
	count := 0
	x := 0.0
	y := 0.0

	for _, name := range names {
		point, ok := trainingPosePointByName(person, name)
		if !ok {
			continue
		}

		count++
		x += point.x
		y += point.y
	}

	if count == 0 {
		return trainingPosePoint{}, false
	}

	return trainingPosePoint{
		x: x / float64(count),
		y: y / float64(count),
	}, true
}

func trainingPosePointDistance(a trainingPosePoint, okA bool, b trainingPosePoint, okB bool) float64 {
	if !okA || !okB {
		return 0
	}

	dx := a.x - b.x
	dy := a.y - b.y
	return math.Sqrt(dx*dx + dy*dy)
}

func trainingPosePersonGeometryFor(person TrainingPosePerson) trainingPosePersonGeometry {
	box, hasBox := trainingNormalizedBox(person.Box)
	center := trainingPosePoint{x: 0.5, y: 0.5}
	if hasBox {
		if x, y, ok := trainingBoxCenter(box); ok {
			center = trainingPosePoint{x: x, y: y}
		}
	}

	leftHip, okLeftHip := trainingPosePointByName(person, "left_hip")
	rightHip, okRightHip := trainingPosePointByName(person, "right_hip")
	leftKnee, okLeftKnee := trainingPosePointByName(person, "left_knee")
	rightKnee, okRightKnee := trainingPosePointByName(person, "right_knee")

	hip, okHip := trainingPoseMidpoint(person, "left_hip", "right_hip")
	shoulder, okShoulder := trainingPoseMidpoint(person, "left_shoulder", "right_shoulder")
	knee, okKnee := trainingPoseMidpoint(person, "left_knee", "right_knee")

	if okHip {
		center = hip
	}

	torsoDX := 0.0
	torsoDY := 0.0
	torsoLen := 0.0
	if okHip && okShoulder {
		torsoDX = math.Abs(hip.x - shoulder.x)
		torsoDY = math.Abs(hip.y - shoulder.y)
		torsoLen = math.Sqrt(torsoDX*torsoDX + torsoDY*torsoDY)
	}

	hipWidth := trainingPosePointDistance(leftHip, okLeftHip, rightHip, okRightHip)
	kneeWidth := trainingPosePointDistance(leftKnee, okLeftKnee, rightKnee, okRightKnee)

	kneesBelowHips := okKnee && okHip && knee.y > hip.y+0.045
	kneesWide := kneeWidth > 0 && kneeWidth >= math.Max(0.11, hipWidth*1.12)
	straddling := kneesBelowHips && kneesWide

	torsoHorizontal := torsoLen >= 0.07 && torsoDX >= torsoDY*1.15
	torsoVertical := torsoLen >= 0.07 && torsoDY >= torsoDX*1.15

	boxHorizontal := hasBox && box.W >= box.H*1.05
	boxVertical := hasBox && box.H >= box.W*1.25

	lying := torsoHorizontal || boxHorizontal
	upright := torsoVertical || boxVertical
	allFours := torsoHorizontal && kneesBelowHips
	bentOrKneeling := allFours || (kneesBelowHips && !straddling)

	return trainingPosePersonGeometry{
		box:            box,
		hasBox:         hasBox,
		center:         center,
		lying:          lying,
		upright:        upright,
		straddling:     straddling,
		kneesBelowHips: kneesBelowHips,
		kneesWide:      kneesWide,
		allFours:       allFours,
		bentOrKneeling: bentOrKneeling,
	}
}

func trainingAddPosePairGeometryScores(
	scores map[string]float64,
	positionSet map[string]bool,
	pose TrainingPosePrediction,
) {
	add := func(label string, score float64) {
		trainingCombinePositionScore(scores, positionSet, label, score)
	}

	pose = trainingReliablePosePrediction(pose)
	if len(pose.Persons) < 2 {
		return
	}

	geometries := make([]trainingPosePersonGeometry, 0, len(pose.Persons))
	for _, person := range pose.Persons {
		geometries = append(geometries, trainingPosePersonGeometryFor(person))
	}

	for i := 0; i < len(geometries); i++ {
		left := geometries[i]
		if !left.hasBox {
			continue
		}

		for j := i + 1; j < len(geometries); j++ {
			right := geometries[j]
			if !right.hasBox {
				continue
			}

			gap := trainingBoxGap(left.box, right.box)
			overlap := trainingBoxOverlapRatio(left.box, right.box)
			horizontalOverlap := trainingBoxHorizontalOverlapRatio(left.box, right.box)
			close := gap <= 0.12 || overlap >= 0.08
			if !close {
				continue
			}

			top := left
			bottom := right
			if right.center.y < left.center.y {
				top = right
				bottom = left
			}

			topAbove := top.center.y <= bottom.center.y-0.055
			strongStack := horizontalOverlap >= 0.35 && (topAbove || overlap >= 0.22)
			topHasRiderShape := top.straddling ||
				top.kneesWide ||
				(top.upright && top.kneesBelowHips)

			if strongStack && topHasRiderShape {
				add("cowgirl", 0.20)
				add("reverse_cowgirl", 0.17)

				if bottom.lying {
					add("cowgirl", 0.12)
					add("reverse_cowgirl", 0.10)
				}
				if top.straddling {
					add("cowgirl", 0.08)
					add("reverse_cowgirl", 0.06)
				}
			}

			if strongStack && bottom.lying {
				add("missionary", 0.14)
				if !top.straddling {
					add("missionary", 0.08)
				}
			}

			bothLying := left.lying && right.lying
			sameLevel := math.Abs(left.center.y-right.center.y) <= 0.15
			sideBySide := math.Abs(left.center.x-right.center.x) >= 0.10
			if bothLying && (sameLevel || sideBySide) {
				add("spooning", 0.18)
				if overlap >= 0.10 {
					add("prone_bone", 0.07)
				}
			}

			leftBent := left.allFours || left.bentOrKneeling
			rightBent := right.allFours || right.bentOrKneeling
			if leftBent != rightBent {
				other := right
				if rightBent {
					other = left
				}

				add("doggy", 0.16)
				if other.upright {
					add("doggy", 0.06)
					add("standing_doggy", 0.07)
				}
				if left.lying || right.lying {
					add("prone_bone", 0.06)
				}
			} else if left.allFours || right.allFours {
				add("doggy", 0.13)
				if left.lying || right.lying {
					add("prone_bone", 0.07)
				}
			}
		}
	}
}

func trainingAddPoseSceneContextScores(
	scores map[string]float64,
	positionSet map[string]bool,
	pred TrainingPrediction,
	pose TrainingPosePrediction,
) {
	add := func(label string, score float64) {
		trainingCombinePositionScore(scores, positionSet, label, score)
	}

	pose = trainingReliablePosePrediction(pose)
	personBoxes := trainingScenePersonBoxes(pred, pose)
	personCount := len(personBoxes)
	if len(pose.Persons) > personCount {
		personCount = len(pose.Persons)
	}

	pair := trainingScenePersonPairSignals(personBoxes)
	trainingAddPosePairGeometryScores(scores, positionSet, pose)

	penisBoxes := trainingBoxesByLabel(pred.Boxes, "penis")
	pussyBoxes := trainingBoxesByLabel(pred.Boxes, "pussy", "vagina", "vulva", "labia")
	assBoxes := trainingBoxesByLabel(pred.Boxes, "ass", "anus")
	breastBoxes := trainingBoxesByLabel(pred.Boxes, "breasts")
	tongueBoxes := trainingBoxesByLabel(pred.Boxes, "tongue")
	toyBoxes := trainingBoxesByLabel(pred.Boxes, "dildo", "vibrator", "strapon", "buttplug")

	hasPenis := len(penisBoxes) > 0
	hasPussy := len(pussyBoxes) > 0
	hasAss := len(assBoxes) > 0
	hasBreasts := len(breastBoxes) > 0
	hasTongue := len(tongueBoxes) > 0
	hasToy := len(toyBoxes) > 0

	headNames := []string{"nose", "left_eye", "right_eye", "left_ear", "right_ear"}
	handNames := []string{"left_wrist", "right_wrist"}
	hipNames := []string{"left_hip", "right_hip"}

	headNearPenis := trainingAnyPoseKeypointNearBoxes(pose, headNames, penisBoxes, 0.09)
	headNearPussy := trainingAnyPoseKeypointNearBoxes(pose, headNames, pussyBoxes, 0.09)
	headNearAss := trainingAnyPoseKeypointNearBoxes(pose, headNames, assBoxes, 0.09)
	handNearPenis := trainingAnyPoseKeypointNearBoxes(pose, handNames, penisBoxes, 0.08)
	handNearPussy := trainingAnyPoseKeypointNearBoxes(pose, handNames, pussyBoxes, 0.08)
	handNearToy := trainingAnyPoseKeypointNearBoxes(pose, handNames, toyBoxes, 0.08)
	hipsNearGenitals := trainingAnyPoseKeypointNearBoxes(pose, hipNames, append(penisBoxes, pussyBoxes...), 0.08)

	if personCount >= 2 {
		add("missionary", 0.04)
		add("doggy", 0.04)
		add("cowgirl", 0.04)
		add("reverse_cowgirl", 0.04)
		add("standing_doggy", 0.04)
		add("spooning", 0.04)
		add("69", 0.03)
	}

	if pair.close {
		add("missionary", 0.04)
		add("doggy", 0.04)
		add("cowgirl", 0.04)
		add("spooning", 0.04)
	}
	if pair.overlap {
		add("missionary", 0.05)
		add("cowgirl", 0.05)
		add("prone_bone", 0.04)
	}
	if pair.horizontal {
		add("spooning", 0.12)
		add("prone_bone", 0.07)
	}
	if pair.vertical {
		add("standing", 0.08)
		add("standing_doggy", 0.10)
	}
	if pair.stacked {
		add("missionary", 0.07)
		add("cowgirl", 0.07)
		add("reverse_cowgirl", 0.06)
		add("facesitting", 0.05)
	}

	if hasPenis && hasPussy {
		add("missionary", 0.08)
		add("doggy", 0.08)
		add("cowgirl", 0.08)
		add("reverse_cowgirl", 0.07)
		add("prone_bone", 0.06)
		add("standing_doggy", 0.06)
		add("spooning", 0.05)

		if trainingAnyBoxesNear(penisBoxes, pussyBoxes, 0.09) || hipsNearGenitals {
			add("missionary", 0.08)
			add("doggy", 0.08)
			add("cowgirl", 0.08)
			add("reverse_cowgirl", 0.07)
		}
	}

	if hasPenis && (headNearPenis || hasTongue) {
		add("blowjob", 0.16)
		if headNearPenis {
			add("blowjob", 0.10)
		}
	}
	if hasPussy && (headNearPussy || hasTongue || trainingAnyBoxesNear(tongueBoxes, pussyBoxes, 0.08)) {
		add("cunnilingus", 0.16)
		if headNearPussy || trainingAnyBoxesNear(tongueBoxes, pussyBoxes, 0.08) {
			add("cunnilingus", 0.10)
		}
	}
	if hasPenis && handNearPenis {
		add("handjob", 0.18)
	}
	if hasPussy && handNearPussy {
		add("fingering", 0.18)
	}
	if hasPenis && hasBreasts && trainingAnyBoxesNear(penisBoxes, breastBoxes, 0.10) {
		add("boobjob", 0.20)
	}
	if hasToy {
		add("toy_play", 0.12)
		if handNearToy ||
			trainingAnyBoxesNear(toyBoxes, pussyBoxes, 0.10) ||
			trainingAnyBoxesNear(toyBoxes, penisBoxes, 0.10) ||
			trainingAnyBoxesNear(toyBoxes, assBoxes, 0.10) {
			add("toy_play", 0.12)
		}
	}
	if personCount >= 2 && (headNearAss || headNearPussy) {
		if hasAss {
			add("facesitting", 0.14)
		}
		if hasPussy && hasPenis {
			add("69", 0.10)
		}
	}
	if hasAss && hasPussy && pair.horizontal {
		add("doggy", 0.07)
		add("prone_bone", 0.07)
	}
}

func trainingApplyPoseToPrediction(pred TrainingPrediction, pose TrainingPosePrediction) TrainingPrediction {
	if pose.Available {
		pred.ModelAvailable = true
	}

	positionSet := stringSet(defaultTrainingLabelsFromJSON().SexPositions)
	poseScores := map[string]float64{}
	contextScores := map[string]float64{}
	reliablePose := TrainingPosePrediction{
		Available: pose.Available,
		Source:    pose.Source,
		Persons:   []TrainingPosePerson{},
	}

	if pose.Available && len(pose.Persons) > 0 {
		pose = trainingAnnotatePosePrediction(pose)
		pred.Persons = pose.Persons
		reliablePose = trainingReliablePosePrediction(pose)
	}

	for _, person := range reliablePose.Persons {
		label := normalizeSexPositionLabel(person.Label)
		if isNoSexPositionLabel(label) || !positionSet[label] {
			continue
		}

		trainingCombinePositionScore(poseScores, positionSet, label, person.Score)

		if quality := trainingPoseKeypointQuality(person); quality > 0 {
			trainingCombinePositionScore(poseScores, positionSet, label, 0.04*quality)
		}
	}

	trainingAddPoseSceneContextScores(contextScores, positionSet, pred, reliablePose)

	bestPosition, bestPositionScore, hasPoseSignal, hasContextSignal :=
		trainingFuseHybridPositionScores(poseScores, contextScores)

	if bestPosition != "" && (hasPoseSignal || bestPositionScore >= trainingPositionContextMinScore) {
		pred.SexPosition = bestPosition
		pred.SexPositionScore = clamp01(bestPositionScore)
	}

	if pose.Available {
		trainingAppendPredictionSource(&pred, "yolo_pose")
	}

	if hasContextSignal {
		trainingAppendPredictionSource(&pred, "box_context")
	}

	return pred
}

func trainingDetectorLabelContent(
	boxes []TrainingBox,
	classMap map[string]int,
	allowEmpty bool,
) ([]byte, error) {
	var lines []string

	for _, box := range boxes {
		label := strings.TrimSpace(box.Label)
		if label == "" {
			continue
		}

		classID, ok := classMap[label]
		if !ok {
			continue
		}

		x := clamp01(box.X)
		y := clamp01(box.Y)
		w := clamp01(box.W)
		h := clamp01(box.H)

		if w <= 0 || h <= 0 {
			continue
		}

		// Frontend/Predictor nutzen x/y als linke obere Ecke.
		// YOLO erwartet x_center/y_center.
		xCenter := clamp01(x + w/2)
		yCenter := clamp01(y + h/2)

		lines = append(lines, fmt.Sprintf(
			"%d %.6f %.6f %.6f %.6f",
			classID,
			xCenter,
			yCenter,
			w,
			h,
		))
	}

	if len(lines) == 0 {
		if allowEmpty {
			return []byte{}, nil
		}
		return nil, errors.New("no valid detector boxes")
	}

	return []byte(strings.Join(lines, "\n") + "\n"), nil
}

func trainingWriteDetectorSample(
	root string,
	sample *TrainingSample,
	boxes []TrainingBox,
	allowEmpty bool,
) error {
	if sample == nil {
		return errors.New("sample missing")
	}

	classMap, err := trainingDetectorClassMap()
	if err != nil {
		return err
	}

	labelContent, err := trainingDetectorLabelContent(boxes, classMap, allowEmpty)
	if err != nil {
		return err
	}

	srcFrame := filepath.Join(root, "frames", sample.SampleID+".jpg")
	if _, err := os.Stat(srcFrame); err != nil {
		return appErrorf("frame missing: %w", err)
	}

	// Stabiler 80/20 Split: gleicher sampleID landet immer im gleichen Split.
	split := trainingStableSplit(sample.SampleID)

	imgDir := filepath.Join(root, "detector", "dataset", "images", split)
	lblDir := filepath.Join(root, "detector", "dataset", "labels", split)

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

	dstFrame := filepath.Join(imgDir, sample.SampleID+".jpg")
	if err := copyFile(srcFrame, dstFrame); err != nil {
		return err
	}

	labelPath := filepath.Join(lblDir, sample.SampleID+".txt")
	return os.WriteFile(labelPath, labelContent, 0644)
}

func trainingSexPositionForFeedback(sample *TrainingSample, req TrainingFeedbackRequest) string {
	if req.Negative {
		return ""
	}

	if req.Correction != nil {
		return normalizeSexPositionLabel(req.Correction.SexPosition)
	}

	if req.Accepted && sample != nil {
		return normalizeSexPositionLabel(sample.Prediction.SexPosition)
	}

	return ""
}

func trainingPosePersonsForSample(root string, sample *TrainingSample) []TrainingPosePerson {
	if sample == nil {
		return nil
	}

	if len(sample.Prediction.Persons) > 0 {
		return sample.Prediction.Persons
	}

	framePath := filepath.Join(root, "frames", sample.SampleID+".jpg")
	if !fileExistsNonEmpty(framePath) {
		return nil
	}

	pose := trainingPredictPose(root, framePath)
	if !pose.Available || len(pose.Persons) == 0 {
		return nil
	}

	return pose.Persons
}

func trainingPoseContextPersonBoxes(boxes []TrainingBox) []TrainingBox {
	out := []TrainingBox{}

	for _, box := range boxes {
		if !trainingIsPersonLikeLabel(box.Label) {
			continue
		}

		if normalized, ok := trainingNormalizedBox(box); ok {
			out = append(out, normalized)
		}
	}

	return out
}

func trainingFilterPosePersonsByContext(
	persons []TrainingPosePerson,
	contextBoxes []TrainingBox,
) []TrainingPosePerson {
	if len(persons) == 0 {
		return persons
	}

	personBoxes := trainingPoseContextPersonBoxes(contextBoxes)
	if len(personBoxes) == 0 {
		return persons
	}

	filtered := []TrainingPosePerson{}

	for _, person := range persons {
		personBox, ok := trainingNormalizedBox(person.Box)
		if !ok {
			continue
		}

		for _, contextBox := range personBoxes {
			if trainingBoxOverlapRatio(personBox, contextBox) >= 0.12 ||
				trainingBoxGap(personBox, contextBox) <= 0.08 {
				filtered = append(filtered, person)
				break
			}
		}
	}

	if len(filtered) == 0 {
		return persons
	}

	return filtered
}

func trainingPoseLabelContent(persons []TrainingPosePerson, classID int) ([]byte, error) {
	lines := []string{}

	for _, person := range persons {
		person = trainingAnnotatePosePersonQuality(person)
		if !trainingPosePersonReliable(person) {
			continue
		}

		if len(person.Keypoints) < trainingPoseKeypointCount {
			continue
		}

		x := clamp01(person.Box.X)
		y := clamp01(person.Box.Y)
		w := clamp01(person.Box.W)
		h := clamp01(person.Box.H)

		if w <= 0 || h <= 0 {
			continue
		}

		xCenter := clamp01(x + w/2)
		yCenter := clamp01(y + h/2)

		parts := []string{
			strconv.Itoa(classID),
			fmt.Sprintf("%.6f", xCenter),
			fmt.Sprintf("%.6f", yCenter),
			fmt.Sprintf("%.6f", w),
			fmt.Sprintf("%.6f", h),
		}

		visible := 0
		for i := 0; i < trainingPoseKeypointCount; i++ {
			kp := person.Keypoints[i]
			kx := clamp01(kp.X)
			ky := clamp01(kp.Y)
			visibility := 0

			if kp.Conf >= trainingPoseKeypointMinConfidence && kx > 0 && ky > 0 {
				visibility = 2
				visible++
			} else {
				kx = 0
				ky = 0
			}

			parts = append(
				parts,
				fmt.Sprintf("%.6f", kx),
				fmt.Sprintf("%.6f", ky),
				strconv.Itoa(visibility),
			)
		}

		if visible < 5 {
			continue
		}

		lines = append(lines, strings.Join(parts, " "))
	}

	if len(lines) == 0 {
		return nil, errors.New("no valid pose persons")
	}

	return []byte(strings.Join(lines, "\n") + "\n"), nil
}

func trainingWritePoseSample(
	root string,
	sample *TrainingSample,
	sexPosition string,
	contextBoxes []TrainingBox,
) error {
	if sample == nil {
		return errors.New("sample missing")
	}

	sexPosition = strings.TrimSpace(sexPosition)
	if isNoSexPositionLabel(sexPosition) {
		return errors.New("pose sex position missing")
	}

	classMap, err := trainingPoseClassMap()
	if err != nil {
		return err
	}

	classID, ok := classMap[sexPosition]
	if !ok {
		return fmt.Errorf("pose class not configured: %s", sexPosition)
	}

	persons := trainingPosePersonsForSample(root, sample)
	persons = trainingFilterPosePersonsByContext(persons, contextBoxes)
	labelContent, err := trainingPoseLabelContent(persons, classID)
	if err != nil {
		return err
	}

	srcFrame := filepath.Join(root, "frames", sample.SampleID+".jpg")
	if _, err := os.Stat(srcFrame); err != nil {
		return appErrorf("frame missing: %w", err)
	}

	split := trainingStableSplit(sample.SampleID)

	imgDir := filepath.Join(root, "pose", "dataset", "images", split)
	lblDir := filepath.Join(root, "pose", "dataset", "labels", split)

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

	dstFrame := filepath.Join(imgDir, sample.SampleID+".jpg")
	if err := copyFile(srcFrame, dstFrame); err != nil {
		return err
	}

	labelPath := filepath.Join(lblDir, sample.SampleID+".txt")
	return os.WriteFile(labelPath, labelContent, 0644)
}

func trainingDeleteDetectorSample(root string, sampleID string) {
	sampleID = strings.TrimSpace(sampleID)
	if sampleID == "" || strings.Contains(sampleID, "/") || strings.Contains(sampleID, "\\") {
		return
	}

	for _, split := range []string{"train", "val"} {
		labelsDir := filepath.Join(root, "detector", "dataset", "labels", split)
		imagesDir := filepath.Join(root, "detector", "dataset", "images", split)

		_ = os.Remove(filepath.Join(labelsDir, sampleID+".txt"))

		for _, ext := range []string{".jpg", ".jpeg", ".png", ".webp"} {
			_ = os.Remove(filepath.Join(imagesDir, sampleID+ext))
		}
	}
}

func trainingDeletePoseSample(root string, sampleID string) {
	sampleID = strings.TrimSpace(sampleID)
	if sampleID == "" || strings.Contains(sampleID, "/") || strings.Contains(sampleID, "\\") {
		return
	}

	for _, split := range []string{"train", "val"} {
		labelsDir := filepath.Join(root, "pose", "dataset", "labels", split)
		imagesDir := filepath.Join(root, "pose", "dataset", "images", split)

		_ = os.Remove(filepath.Join(labelsDir, sampleID+".txt"))

		for _, ext := range []string{".jpg", ".jpeg", ".png", ".webp"} {
			_ = os.Remove(filepath.Join(imagesDir, sampleID+ext))
		}
	}
}

func trainingSyncPoseDataset(root string) (int, error) {
	if err := trainingEnsurePoseDirs(root); err != nil {
		return 0, err
	}

	items, err := trainingReadAnnotations(root)
	if err != nil {
		return 0, err
	}

	written := 0

	for _, item := range items {
		sampleID := strings.TrimSpace(item.SampleID)
		if sampleID == "" {
			continue
		}

		sample := &TrainingSample{
			SampleID:         item.SampleID,
			FrameURL:         item.FrameURL,
			SourceFile:       item.SourceFile,
			SourcePath:       item.SourcePath,
			SourceSizeBytes:  item.SourceSizeBytes,
			Second:           item.Second,
			CreatedAt:        item.CreatedAt,
			UncertaintyScore: 0,
			Prediction:       item.Prediction,
		}

		effective := trainingEffectiveCorrection(item)
		sexPosition := strings.TrimSpace(effective.SexPosition)

		trainingDeletePoseSample(root, sampleID)

		if item.Negative || isNoSexPositionLabel(sexPosition) {
			continue
		}

		if err := trainingWritePoseSample(root, sample, sexPosition, effective.Boxes); err != nil {
			appLogln("pose sample sync skipped:", sampleID, err)
			continue
		}

		written++
	}

	return written, nil
}

func trainingEnsureDetectorValidationSample(root string) error {
	trainImages := filepath.Join(root, "detector", "dataset", "images", "train")
	trainLabels := filepath.Join(root, "detector", "dataset", "labels", "train")
	valImages := filepath.Join(root, "detector", "dataset", "images", "val")
	valLabels := filepath.Join(root, "detector", "dataset", "labels", "val")

	currentVal := trainingCountDetectorSamples(valImages, valLabels)
	currentPositiveVal := trainingCountPositiveDetectorSamples(valImages, valLabels)
	if currentVal >= minDetectorValCount && currentPositiveVal > 0 {
		return nil
	}

	if trainingCountDetectorSamples(trainImages, trainLabels) < minDetectorTrainCount ||
		trainingCountPositiveDetectorSamples(trainImages, trainLabels) == 0 {
		return nil
	}

	entries, err := os.ReadDir(trainImages)
	if err != nil {
		return nil
	}

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

	copied := 0
	needed := max(0, minDetectorValCount-currentVal)
	needsPositive := currentPositiveVal == 0

	sort.SliceStable(entries, func(i, j int) bool {
		iID := strings.TrimSuffix(entries[i].Name(), filepath.Ext(entries[i].Name()))
		jID := strings.TrimSuffix(entries[j].Name(), filepath.Ext(entries[j].Name()))

		return fileExistsNonEmpty(filepath.Join(trainLabels, iID+".txt")) &&
			!fileExistsNonEmpty(filepath.Join(trainLabels, jID+".txt"))
	})

	for _, e := range entries {
		if copied >= needed && !needsPositive {
			break
		}

		if e.IsDir() {
			continue
		}

		ext := strings.ToLower(filepath.Ext(e.Name()))
		if ext != ".jpg" && ext != ".jpeg" && ext != ".png" && ext != ".webp" {
			continue
		}

		id := strings.TrimSuffix(e.Name(), filepath.Ext(e.Name()))

		srcImage := filepath.Join(trainImages, e.Name())
		srcLabel := filepath.Join(trainLabels, id+".txt")

		if !fileExistsNonEmpty(srcImage) || !fileExists(srcLabel) {
			continue
		}

		dstImage := filepath.Join(valImages, e.Name())
		dstLabel := filepath.Join(valLabels, id+".txt")

		if fileExistsNonEmpty(dstImage) && fileExists(dstLabel) {
			continue
		}

		if err := copyFile(srcImage, dstImage); err != nil {
			return err
		}
		if err := copyFile(srcLabel, dstLabel); err != nil {
			return err
		}

		copied++
		if fileExistsNonEmpty(srcLabel) {
			needsPositive = false
		}
	}

	return nil
}

func trainingEnsurePoseValidationSample(root string) error {
	trainImages := filepath.Join(root, "pose", "dataset", "images", "train")
	trainLabels := filepath.Join(root, "pose", "dataset", "labels", "train")
	valImages := filepath.Join(root, "pose", "dataset", "images", "val")
	valLabels := filepath.Join(root, "pose", "dataset", "labels", "val")

	currentVal := trainingCountDetectorSamples(valImages, valLabels)
	if currentVal >= minPoseValCount {
		return nil
	}

	if trainingCountDetectorSamples(trainImages, trainLabels) < minPoseTrainCount {
		return nil
	}

	entries, err := os.ReadDir(trainImages)
	if err != nil {
		return nil
	}

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

	copied := 0
	needed := max(0, minPoseValCount-currentVal)

	for _, e := range entries {
		if copied >= needed {
			break
		}

		if e.IsDir() {
			continue
		}

		ext := strings.ToLower(filepath.Ext(e.Name()))
		if ext != ".jpg" && ext != ".jpeg" && ext != ".png" && ext != ".webp" {
			continue
		}

		id := strings.TrimSuffix(e.Name(), filepath.Ext(e.Name()))

		srcImage := filepath.Join(trainImages, e.Name())
		srcLabel := filepath.Join(trainLabels, id+".txt")

		if !fileExistsNonEmpty(srcImage) || !fileExistsNonEmpty(srcLabel) {
			continue
		}

		dstImage := filepath.Join(valImages, e.Name())
		dstLabel := filepath.Join(valLabels, id+".txt")

		if fileExistsNonEmpty(dstImage) && fileExistsNonEmpty(dstLabel) {
			continue
		}

		if err := copyFile(srcImage, dstImage); err != nil {
			return err
		}
		if err := copyFile(srcLabel, dstLabel); err != nil {
			return err
		}

		copied++
	}

	return nil
}

func trainingStableSplit(sampleID string) string {
	sum := sha1.Sum([]byte(sampleID))
	if int(sum[0])%5 == 0 {
		return "val"
	}
	return "train"
}

func copyFile(src string, dst string) error {
	b, err := os.ReadFile(src)
	if err != nil {
		return err
	}
	return os.WriteFile(dst, b, 0644)
}

func trainingEmptyPrediction(source string) TrainingPrediction {
	return TrainingPrediction{
		ModelAvailable:   false,
		Source:           source,
		SexPosition:      trainingNoSexPositionLabel,
		SexPositionScore: 0,
		PeoplePresent:    []TrainingScoredLabel{},
		BodyPartsPresent: []TrainingScoredLabel{},
		ObjectsPresent:   []TrainingScoredLabel{},
		ClothingPresent:  []TrainingScoredLabel{},
		Boxes:            []TrainingBox{},
	}
}

func trainingPythonExe() string {
	v := strings.TrimSpace(os.Getenv("TRAINING_PYTHON"))
	if v != "" {
		return v
	}
	return "python"
}

func trainingProjectRoot() string {
	wd, err := os.Getwd()
	if err != nil {
		return "."
	}

	if _, err := os.Stat(filepath.Join(wd, "backend", "ml", "predict_detector_model.py")); err == nil {
		return wd
	}

	if _, err := os.Stat(filepath.Join(wd, "ml", "predict_detector_model.py")); err == nil {
		return filepath.Dir(wd)
	}

	parent := filepath.Dir(wd)
	if _, err := os.Stat(filepath.Join(parent, "backend", "ml", "predict_detector_model.py")); err == nil {
		return parent
	}

	return wd
}

func trainingScriptPath(name string) string {
	// 1) Eingebettete Scripts bevorzugen.
	if dir, err := trainingEmbeddedMLDir(); err == nil {
		p := filepath.Join(dir, name)
		if _, err := os.Stat(p); err == nil {
			return p
		}
	}

	// 2) App-/Backend-relativ wie record_paths.go.
	if backendRoot, err := trainingBackendRootDir(); err == nil {
		p := filepath.Join(backendRoot, "ml", name)
		if _, err := os.Stat(p); err == nil {
			return p
		}
	}

	// 3) Dev-Fallback.
	root := trainingProjectRoot()

	p := filepath.Join(root, "backend", "ml", name)
	if _, err := os.Stat(p); err == nil {
		return p
	}

	p = filepath.Join("ml", name)
	if _, err := os.Stat(p); err == nil {
		return p
	}

	return filepath.Join(root, "backend", "ml", name)
}

func isTempBuildDir(dir string) bool {
	low := strings.ToLower(filepath.Clean(dir))

	return strings.Contains(low, `\appdata\local\temp`) ||
		strings.Contains(low, `\temp\`) ||
		strings.Contains(low, `\tmp\`) ||
		strings.Contains(low, `\go-build`) ||
		strings.Contains(low, `/tmp/`) ||
		strings.Contains(low, `/go-build`)
}

func trainingBackendRootDir() (string, error) {
	if script, err := resolvePathRelativeToApp(filepath.Join("ml", "predict_detector_model.py")); err == nil {
		if st, statErr := os.Stat(script); statErr == nil && !st.IsDir() {
			return filepath.Dir(filepath.Dir(script)), nil
		}
	}

	if script, err := resolvePathRelativeToApp(filepath.Join("backend", "ml", "predict_detector_model.py")); err == nil {
		if st, statErr := os.Stat(script); statErr == nil && !st.IsDir() {
			return filepath.Dir(filepath.Dir(script)), nil
		}
	}

	if dir, err := exeDir(); err == nil && strings.TrimSpace(dir) != "" && !isTempBuildDir(dir) {
		return dir, nil
	}

	wd, err := os.Getwd()
	if err != nil {
		return "", err
	}

	if _, err := os.Stat(filepath.Join(wd, "ml", "predict_detector_model.py")); err == nil {
		return wd, nil
	}

	if _, err := os.Stat(filepath.Join(wd, "backend", "ml", "predict_detector_model.py")); err == nil {
		return filepath.Join(wd, "backend"), nil
	}

	projectRoot := trainingProjectRoot()
	return filepath.Join(projectRoot, "backend"), nil
}

func trainingRootDir() (string, error) {
	// Optionaler Override, falls du später explizit einen anderen Speicherort willst.
	// Relative Pfade werden wie in record_paths.go app-relativ aufgelöst.
	if override := strings.TrimSpace(os.Getenv("TRAINING_ROOT")); override != "" {
		root, err := resolvePathRelativeToApp(override)
		if err != nil {
			return "", err
		}

		root, err = filepath.Abs(root)
		if err != nil {
			return "", err
		}

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

		return root, nil
	}

	backendRoot, err := trainingBackendRootDir()
	if err != nil {
		return "", err
	}

	root, err := filepath.Abs(filepath.Join(backendRoot, "generated", "training"))
	if err != nil {
		return "", err
	}

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

	return root, nil
}

func trainingWriteSample(root string, sample *TrainingSample) error {
	path := filepath.Join(root, "samples", sample.SampleID+".json")
	b, err := json.MarshalIndent(sample, "", "  ")
	if err != nil {
		return err
	}
	return os.WriteFile(path, b, 0644)
}

func trainingReadSample(root string, sampleID string) (*TrainingSample, error) {
	if strings.Contains(sampleID, "/") || strings.Contains(sampleID, "\\") {
		return nil, errors.New("invalid sample id")
	}

	path := filepath.Join(root, "samples", sampleID+".json")

	b, err := os.ReadFile(path)
	if err != nil {
		return nil, err
	}

	var sample TrainingSample
	if err := json.Unmarshal(b, &sample); err != nil {
		return nil, err
	}

	return &sample, nil
}

func trainingAppendAnnotation(root string, annotation TrainingAnnotation) error {
	path := filepath.Join(root, "feedback.jsonl")

	f, err := os.OpenFile(path, os.O_CREATE|os.O_APPEND|os.O_WRONLY, 0644)
	if err != nil {
		return err
	}
	defer f.Close()

	b, err := json.Marshal(annotation)
	if err != nil {
		return err
	}

	if _, err := f.Write(append(b, '\n')); err != nil {
		return err
	}

	return nil
}

func trainingWriteAnnotations(root string, items []TrainingAnnotation) error {
	path := filepath.Join(root, "feedback.jsonl")
	tmpPath := path + ".tmp"

	var b strings.Builder

	for _, item := range items {
		line, err := json.Marshal(item)
		if err != nil {
			return err
		}

		b.Write(line)
		b.WriteByte('\n')
	}

	if err := os.WriteFile(tmpPath, []byte(b.String()), 0644); err != nil {
		return err
	}

	return os.Rename(tmpPath, path)
}

func trainingCountAnnotations(path string) (int, error) {
	b, err := os.ReadFile(path)
	if err != nil {
		return 0, err
	}

	text := strings.TrimSpace(string(b))
	if text == "" {
		return 0, nil
	}

	return len(strings.Split(text, "\n")), nil
}

func trainingProbeDurationSeconds(videoPath string) float64 {
	settings := getSettings()

	ffmpeg := strings.TrimSpace(settings.FFmpegPath)
	ffprobe := "ffprobe"

	if ffmpeg != "" {
		dir := filepath.Dir(ffmpeg)
		base := filepath.Base(ffmpeg)
		if strings.Contains(strings.ToLower(base), "ffmpeg") {
			ffprobeBase := strings.Replace(base, "ffmpeg", "ffprobe", 1)
			ffprobe = filepath.Join(dir, ffprobeBase)
		}
	}

	cmd := exec.Command(
		ffprobe,
		"-v", "error",
		"-show_entries", "format=duration",
		"-of", "default=noprint_wrappers=1:nokey=1",
		videoPath,
	)

	cmd.SysProcAttr = &syscall.SysProcAttr{
		HideWindow:    true,
		CreationFlags: 0x08000000,
	}

	out, err := cmd.Output()
	if err != nil {
		return 0
	}

	v, err := strconv.ParseFloat(strings.TrimSpace(string(out)), 64)
	if err != nil || math.IsNaN(v) || math.IsInf(v, 0) || v <= 0 {
		return 0
	}

	return v
}

func trainingRandomSecond(duration float64) float64 {
	if duration <= 2 {
		return 0
	}

	minSec := 1.0
	maxSec := math.Max(minSec, duration-1)

	return minSec + rand.Float64()*(maxSec-minSec)
}

func trainingMakeSampleID(videoPath string, second float64) string {
	h := sha1.New()
	_, _ = h.Write([]byte(videoPath))
	_, _ = h.Write([]byte("|"))
	_, _ = h.Write([]byte(strconv.FormatFloat(second, 'f', 3, 64)))
	_, _ = h.Write([]byte("|"))
	_, _ = h.Write([]byte(strconv.FormatInt(time.Now().UnixNano(), 10)))
	return hex.EncodeToString(h.Sum(nil))[:20]
}

func trainingWriteJSON(w http.ResponseWriter, status int, v any) {
	w.Header().Set("Content-Type", "application/json")
	w.Header().Set("Cache-Control", "no-store")
	w.WriteHeader(status)
	_ = json.NewEncoder(w).Encode(v)
}

func trainingWriteError(w http.ResponseWriter, status int, msg string) {
	trainingWriteJSON(w, status, map[string]any{
		"ok":    false,
		"error": msg,
	})
}