#!/usr/bin/env python3
"""
analyze.py — Predhodna analiza CELEGA videa pred trim-anjem.

Naredi:
1. Whisper transcript celega videa (auto-detect jezika ali user-specified)
2. Energy profile (RMS dB na 1s windows)
3. Structural detection (vocal/instrumental sections, energy peaks)
4. Pametno izbere clip range (lahko >30s, vključi pre-chorus)
5. Detekcija instrumentalnih pesmi (no_subs auto)

Output: JSON s podatki za clip.py
"""

import argparse
import json
import os
import re
import subprocess
import sys
import tempfile
import time
from pathlib import Path


def get_video_duration(path):
    r = subprocess.run(
        ["ffprobe", "-v", "error", "-show_entries", "format=duration",
         "-of", "default=nw=1:nokey=1", str(path)],
        capture_output=True, text=True
    )
    try:
        return float(r.stdout.strip())
    except ValueError:
        return 0.0


def extract_audio(video_path):
    """Extract avdio v 16kHz mono WAV za Whisper + energy."""
    audio = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
    audio.close()
    subprocess.run(
        ["ffmpeg", "-y", "-i", str(video_path), "-vn",
         "-ac", "1", "-ar", "16000", "-c:a", "pcm_s16le", audio.name],
        check=True, capture_output=True
    )
    return audio.name


def detect_language_from_filename(filename_hint):
    """Detektiraj jezik iz imena datoteke na podlagi znanih izvajalcev/besed.
    
    Vrne ISO 639-1 ('sl', 'de', 'en', 'hr'...) ali None.
    """
    if not filename_hint:
        return None
    
    name = filename_hint.lower()
    
    # Slovenski izvajalci (narodno-zabavna, pop, rock)
    SLO_ARTISTS = [
        "avseniki", "avsenik", "modrijani", "veseli dolenjci",
        "čuki", "atomik harmonik", "alfi nipič", "helena blagne",
        "siddharta", "magnifico", "vlado kreslin", "zaklonišče prepeva",
        "perpetuum jazzile", "tabu", "natalija verboten", "klavdija",
        "iztok mlakar", "rok'n'band", "okrog cele zemlje", "ansambel",
        "miran rudan", "andrej šifrer", "mi2", "elvis jackson",
        "tanja žagar", "manca špik", "saša lendero", "rebeka dremelj",
        "nuša derenda", "alenka godec", "prifarski muzikanti",
        "nova generacija", "polka", "narodno-zabavna",
    ]
    SLO_KEYWORDS = ["pazi", "morju", "zveza", "domovina", "ljubim", "srce", "majhna", 
                    "prav", "nazaj", "noč", "dom", "pomoč", "bolha", "preko"]
    
    # Nemški izvajalci (Schlager, Volksmusik)
    DE_ARTISTS = [
        "ben zucker", "andrea berg", "helene fischer", "andreas gabalier",
        "amigos", "kastelruther spatzen", "florian silbereisen", "voxxclub",
        "wolfgang petry", "mickie krause", "die toten hosen", "rammstein",
        "udo lindenberg", "die ärzte", "westernhagen", "peter maffay",
        "matthias reim", "die zillertaler", "die jungen zillertaler",
        "stefan mross", "marianne", "michael wendler", "vincent gross",
        "schlager", "volksmusik",
    ]
    DE_KEYWORDS = ["liebe", "herz", "ohne", "dich", "leben", "nacht", "tag",
                   "schön", "mädchen", "sonne", "himmel", "wenn", "nur",
                   "bist", "hast", "dass", "weiß", "kann", "auch"]
    
    # Hrvaški/srbski izvajalci
    HR_ARTISTS = [
        "thompson", "miroslav škoro", "oliver dragojević", "gibonni",
        "severina", "tony cetinski", "psihomodo pop", "prljavo kazalište",
        "parni valjak", "lepa brena", "ceca", "aca lukas", "mile kitić",
        "halid bešlić", "dino merlin", "zdravko čolić", "magazin",
    ]
    HR_KEYWORDS = ["volim", "ljubav", "srce", "danas", "noćas", "more",
                   "majka", "domovina", "zauvijek", "samo", "ćemo"]
    
    # Angleški izvajalci (preveč jih je za listo, raje preverim ne-SL/DE/HR znake)
    EN_KEYWORDS = ["love", "song", "feat", "remix", "official", "music", "video",
                   "remastered", "lyrics", "by", "with", "tonight", "forever",
                   "heart", "dance", "party", "summer"]
    
    score = {"sl": 0, "de": 0, "hr": 0, "en": 0, "it": 0, "es": 0, "fr": 0}
    
    # Artist matches (težji)
    for a in SLO_ARTISTS:
        if a in name:
            score["sl"] += 5
    for a in DE_ARTISTS:
        if a in name:
            score["de"] += 5
    for a in HR_ARTISTS:
        if a in name:
            score["hr"] += 5
    
    # Keyword matches
    for kw in SLO_KEYWORDS:
        if kw in name.split() or f" {kw} " in f" {name} ":
            score["sl"] += 1
    for kw in DE_KEYWORDS:
        if kw in name.split() or f" {kw} " in f" {name} ":
            score["de"] += 1
    for kw in HR_KEYWORDS:
        if kw in name.split() or f" {kw} " in f" {name} ":
            score["hr"] += 1
    for kw in EN_KEYWORDS:
        if kw in name.split() or f" {kw} " in f" {name} ":
            score["en"] += 1
    
    # Slovenska abeceda (č, ž, š) brez đ (ki je hrvaška)
    if any(c in name for c in "čžš") and "đ" not in name:
        score["sl"] += 2
    # Nemška abeceda (ä ö ü ß)
    if any(c in name for c in "äöüß"):
        score["de"] += 2
    # Hrvaška abeceda (đ)
    if "đ" in name:
        score["hr"] += 2
    
    if not any(score.values()):
        return None
    
    best = max(score.items(), key=lambda x: x[1])
    if best[1] >= 2:  # threshold
        return best[0]
    return None


def transcribe_with_elevenlabs(audio_path, lang=None, model="scribe_v1", filename_hint=None):
    """ElevenLabs Scribe transkripcija (najboljša multilingual accuracy 2026).
    
    lang: ISO 639-1 ('de', 'sl', 'hr') — če None, probamo iz filename_hint
    Pricing: ~$0.40/h (~$0.022 per 200s pesem).
    """
    import urllib.request
    import urllib.error
    import uuid

    api_key = os.environ.get("ELEVENLABS_API_KEY")
    if not api_key:
        print("   ⚠️ ELEVENLABS_API_KEY ni nastavljen", file=sys.stderr)
        return None

    # Auto-detect lang from filename če uporabnik ni eksplicitno izbral
    if not lang and filename_hint:
        guessed = detect_language_from_filename(filename_hint)
        if guessed:
            lang = guessed
            print(f"   🔍 Lang iz filename '{filename_hint}': {lang}", file=sys.stderr)

    # ISO 639-1 → 639-3 mapping (Scribe uses 639-3)
    LANG_1_TO_3 = {
        "en": "eng", "de": "deu", "sl": "slv", "hr": "hrv", "bs": "bos",
        "sr": "srp", "it": "ita", "es": "spa", "fr": "fra", "pt": "por",
        "ru": "rus", "pl": "pol", "cs": "ces", "sk": "slk", "hu": "hun",
        "ro": "ron", "nl": "nld", "sv": "swe", "no": "nor", "da": "dan",
        "fi": "fin", "tr": "tur", "ar": "ara", "uk": "ukr", "bg": "bul",
        "el": "ell", "he": "heb", "ja": "jpn", "ko": "kor", "zh": "zho",
    }
    LANG_3_TO_1 = {v: k for k, v in LANG_1_TO_3.items()}

    # Multipart upload
    boundary = uuid.uuid4().hex
    parts = []

    def add_text(name, value):
        parts.append(
            f"--{boundary}\r\nContent-Disposition: form-data; "
            f"name=\"{name}\"\r\n\r\n{value}\r\n".encode()
        )

    def add_file(name, filename, content, ctype):
        parts.append(
            f"--{boundary}\r\nContent-Disposition: form-data; "
            f"name=\"{name}\"; filename=\"{filename}\"\r\n"
            f"Content-Type: {ctype}\r\n\r\n".encode() + content + b"\r\n"
        )

    with open(audio_path, "rb") as f:
        audio_content = f.read()

    # Limit: ElevenLabs Scribe supports up to ~25 MB / 4.5h per request
    if len(audio_content) > 24 * 1024 * 1024:
        print(f"   ⚠️ Audio {len(audio_content)/1024/1024:.1f} MB > 24 MB limit, fallback", file=sys.stderr)
        return None

    add_text("model_id", model)
    add_text("timestamps_granularity", "word")
    # tag_audio_events=true je kritično: brez tega Scribe predčasno preneha s transkripcijo
    # ko zazna instrumentalni del (npr. polka harmonika prevzame). Z true vstavi oznake
    # kot "(glasba)" in nadaljuje transkripcijo do konca audia.
    # Te oznake potem post-processing odstrani iz besedila.
    add_text("tag_audio_events", "true")
    if lang:
        scribe_lang = LANG_1_TO_3.get(lang, lang)
        add_text("language_code", scribe_lang)
    add_file("file", "audio.mp3", audio_content, "audio/mpeg")
    parts.append(f"--{boundary}--\r\n".encode())
    body = b"".join(parts)

    print(f"   📡 ElevenLabs Scribe ({model}, {len(audio_content)/1024/1024:.1f} MB, "
          f"lang={lang or 'auto'})...", file=sys.stderr)

    req = urllib.request.Request(
        "https://api.elevenlabs.io/v1/speech-to-text",
        data=body,
        headers={
            "xi-api-key": api_key,
            "Content-Type": f"multipart/form-data; boundary={boundary}",
        },
    )

    try:
        with urllib.request.urlopen(req, timeout=300) as resp:
            data = json.loads(resp.read().decode())
    except urllib.error.HTTPError as e:
        body_err = e.read().decode("utf-8", errors="replace")[:500]
        print(f"   ❌ Scribe HTTP {e.code}: {body_err}", file=sys.stderr)
        return None
    except Exception as e:
        print(f"   ❌ Scribe exception: {e}", file=sys.stderr)
        return None

    # Convert response to our standard format
    detected_lang_3 = data.get("language_code", "unknown")
    detected_lang_1 = LANG_3_TO_1.get(detected_lang_3, detected_lang_3[:2])
    detected_prob = data.get("language_probability", 1.0)

    # Scribe returns flat list of words (not segments)
    # We group words into pseudo-segments using **smart phrase-aware segmentation**:
    # - Close on long pause (>= 0.4s) — natural breath/phrase boundary
    # - OR after sentence-ending punctuation (. ! ?)
    # - OR after 4 seconds (max segment length for readable subtitle)
    # This gives ~3-7 word segments matching natural sung phrases.
    words = data.get("words", [])
    segments = []

    if words:
        # Filter out:
        # 1. whitespace tokens
        # 2. audio event tags type='audio_event' or text in (parenthesis) like "(glasba)", "(music)"
        real_words = []
        for w in words:
            t = w.get("text", "").strip()
            wtype = w.get("type", "word")
            # Skip non-word events
            if wtype != "word":
                continue
            if not t:
                continue
            # Skip parenthesized audio events (legacy fallback)
            if t.startswith("(") and t.endswith(")"):
                continue
            real_words.append(w)
        
        if real_words:
            current_seg_words = []
            seg_start = real_words[0].get("start", 0)
            
            for i, w in enumerate(real_words):
                current_seg_words.append(w)
                w_end = w.get("end", w.get("start", 0))
                w_text = w.get("text", "")
                
                close = False
                # Decide if we should close the segment
                if i + 1 < len(real_words):
                    next_start = real_words[i + 1].get("start", w_end)
                    pause = next_start - w_end
                    seg_duration = w_end - seg_start
                    
                    # Trigger close on:
                    # 1. Long pause (>= 0.4s) = phrase boundary
                    # 2. Sentence-ending punctuation
                    # 3. Segment is long enough (>= 4s)
                    if pause >= 0.4:
                        close = True
                    elif seg_duration >= 4.0 and pause >= 0.15:
                        close = True
                    elif w_text.rstrip().endswith(('.', '!', '?')) and pause >= 0.2:
                        close = True
                    elif seg_duration >= 5.5:  # hard cap
                        close = True
                else:
                    close = True  # last word
                
                if close:
                    seg_text = " ".join(ww.get("text", "") for ww in current_seg_words).strip()
                    if seg_text:
                        segments.append({
                            "start": seg_start,
                            "end": w_end,
                            "text": seg_text,
                            "words": [
                                {
                                    "start": ww.get("start", 0),
                                    "end": ww.get("end", 0),
                                    "text": ww.get("text", ""),
                                }
                                for ww in current_seg_words
                            ],
                        })
                    # Reset
                    current_seg_words = []
                    if i + 1 < len(real_words):
                        seg_start = real_words[i + 1].get("start", 0)

    # ── HALLUCINATION DETECTION ──
    # Scribe občasno vrne single dolg segment z 1-2 besedama (10-100s ene besede).
    # To je halucinacija pri instrumentalih.
    hallucination_segs = []
    total_audio_duration = max((s["end"] for s in segments), default=0)
    coverage = 0
    for s in segments:
        seg_dur = s["end"] - s["start"]
        word_count = len(s.get("words", []))
        if seg_dur > 15 and word_count < 5:
            hallucination_segs.append(s)
        else:
            coverage += seg_dur
    
    coverage_pct = coverage / total_audio_duration * 100 if total_audio_duration else 0
    
    if hallucination_segs:
        print(f"   ⚠️ Halucinacija(e) zaznana(e): {len(hallucination_segs)} segment(ov) "
              f"daljših od 15s z manj kot 5 besedami:", file=sys.stderr)
        for h in hallucination_segs:
            print(f"      [{h['start']:.1f}-{h['end']:.1f}s] = {h['end']-h['start']:.0f}s "
                  f"({len(h.get('words', []))} bes.) text={h.get('text', '')[:50]!r}", file=sys.stderr)
        print(f"   📊 Pravo pokritje: {coverage:.1f}s / {total_audio_duration:.1f}s "
              f"= {coverage_pct:.0f}%", file=sys.stderr)
    
    print(f"   ✅ Scribe: {len(words)} words → {len(segments)} segments, "
          f"lang={detected_lang_1} (p={detected_prob:.2f})", file=sys.stderr)

    return {
        "language": detected_lang_1,
        "language_probability": float(detected_prob),
        "segments": segments,
        "_provider": "elevenlabs",
        "_hallucination_count": len(hallucination_segs),
        "_coverage_pct": coverage_pct,
    }


def transcribe_with_gemini(audio_path, lang=None, filename_hint=None):
    """Gemini 3 Pro audio transcription — fallback za narodno-zabavne pesmi
    kjer Scribe halucinarala.
    
    Prednosti:
    - Pravilna besedila slovenskih, hrvaških in drugih "manjšinskih" jezikov
    - Ne halucinira pri instrumentalnih sekcijah
    - Razume kontekst pesmi (lirika)
    
    Slabosti:
    - Počasen (~100s na 2min audio)
    - Dražji ($0.20 vs $0.013)
    - Timestamps včasih off za 1-2s
    """
    import urllib.request
    import urllib.error
    
    api_key = os.environ.get("GEMINI_API_KEY")
    if not api_key:
        print(f"   ❌ Gemini fallback: GEMINI_API_KEY missing", file=sys.stderr)
        return None
    
    print(f"🧠 Gemini 3 Pro transcribing {audio_path}...", file=sys.stderr)
    audio_size_mb = os.path.getsize(audio_path) / 1024 / 1024
    print(f"   📦 Audio size: {audio_size_mb:.1f} MB", file=sys.stderr)
    
    try:
        # 1. Upload audio prek Files API (resumable)
        upload_url_base = "https://generativelanguage.googleapis.com/upload/v1beta/files"
        with open(audio_path, 'rb') as f:
            audio_bytes = f.read()
        
        # Step 1: start
        headers_start = {
            'X-Goog-Upload-Protocol': 'resumable',
            'X-Goog-Upload-Command': 'start',
            'X-Goog-Upload-Header-Content-Length': str(len(audio_bytes)),
            'X-Goog-Upload-Header-Content-Type': 'audio/mp3',
            'Content-Type': 'application/json',
        }
        req_start = urllib.request.Request(
            f"{upload_url_base}?key={api_key}",
            data=json.dumps({"file": {"display_name": "reels_audio"}}).encode(),
            headers=headers_start, method='POST'
        )
        with urllib.request.urlopen(req_start, timeout=30) as resp:
            upload_url = resp.headers.get('X-Goog-Upload-URL')
        
        # Step 2: upload bytes
        headers_upload = {
            'Content-Length': str(len(audio_bytes)),
            'X-Goog-Upload-Offset': '0',
            'X-Goog-Upload-Command': 'upload, finalize',
        }
        req_upload = urllib.request.Request(
            upload_url, data=audio_bytes,
            headers=headers_upload, method='POST'
        )
        with urllib.request.urlopen(req_upload, timeout=120) as resp:
            file_info = json.loads(resp.read().decode())
            file_uri = file_info['file']['uri']
        
        print(f"   ✓ Uploaded to Gemini Files API", file=sys.stderr)
        # Manjši delay da se file procesi
        time.sleep(2)
        
        # 2. Generate transcript
        gen_url = (f"https://generativelanguage.googleapis.com/v1beta/"
                   f"models/gemini-3-pro-preview:generateContent?key={api_key}")
        
        lang_hint = ""
        if filename_hint:
            lang_hint = f"\nFilename hint: {filename_hint}"
        if lang:
            lang_hint += f"\nLanguage: {lang}"
        
        prompt = f"""Transcribe this song with precise word-level timestamps.{lang_hint}

Return ONLY valid JSON in this EXACT format (no markdown fences, no explanation):
{{
  "language": "sl",
  "segments": [
    {{
      "start": 0.5,
      "end": 4.2,
      "text": "Besedilo segmenta",
      "words": [
        {{"start": 0.5, "end": 0.9, "text": "Besedilo"}},
        {{"start": 1.0, "end": 1.4, "text": "segmenta"}}
      ]
    }}
  ]
}}

Rules:
- Only transcribe vocal singing, NOT instrumental sections
- Each segment is a complete musical phrase (typically 2-4 seconds)
- Include word-level timestamps for EVERY word
- Use proper orthography (š, č, ž for Slavic; ä, ö, ü for German etc.)
- Skip instrumental breaks (don't fill with silence segments)
- Be very accurate with timestamps - this is for video subtitle generation
- DO NOT hallucinate words during instrumental sections
- DO NOT include trailing commas in JSON

Output ONLY the JSON object."""
        
        payload = {
            "contents": [{
                "parts": [
                    {"text": prompt},
                    {"file_data": {"mime_type": "audio/mp3", "file_uri": file_uri}}
                ]
            }],
            "generationConfig": {
                "temperature": 0.0,
                "maxOutputTokens": 32000,
            }
        }
        
        req_gen = urllib.request.Request(
            gen_url,
            data=json.dumps(payload).encode(),
            headers={'Content-Type': 'application/json'},
            method='POST'
        )
        
        t0 = time.time()
        with urllib.request.urlopen(req_gen, timeout=300) as resp:
            result = json.loads(resp.read().decode())
        elapsed = time.time() - t0
        
        usage = result.get('usageMetadata', {})
        print(f"   ✓ Gemini 3 Pro response v {elapsed:.0f}s "
              f"(in: {usage.get('promptTokenCount', 0)}, "
              f"out: {usage.get('candidatesTokenCount', 0)}, "
              f"thoughts: {usage.get('thoughtsTokenCount', 0)})", file=sys.stderr)
        
        # 3. Parse JSON output
        candidate_text = result['candidates'][0]['content']['parts'][0]['text'].strip()
        
        # Pobriši markdown code fences če so
        if candidate_text.startswith('```'):
            # ```json\n...\n```
            lines = candidate_text.split('\n')
            if lines[0].startswith('```'):
                lines = lines[1:]
            if lines and lines[-1].rstrip() == '```':
                lines = lines[:-1]
            candidate_text = '\n'.join(lines)
        
        # Try-except za JSON z popravki za pogoste težave
        parsed = None
        try:
            parsed = json.loads(candidate_text)
        except json.JSONDecodeError as e:
            # Trailing comma fix
            import re as _re
            cleaned = _re.sub(r',(\s*[}\]])', r'\1', candidate_text)
            try:
                parsed = json.loads(cleaned)
                print(f"   ✓ Fixed trailing commas in Gemini JSON", file=sys.stderr)
            except json.JSONDecodeError as e2:
                print(f"   ❌ Gemini JSON parse failed: {e2}", file=sys.stderr)
                print(f"   First 500 chars: {candidate_text[:500]}", file=sys.stderr)
                return None
        
        if not parsed or not parsed.get('segments'):
            print(f"   ❌ Gemini returned no segments", file=sys.stderr)
            return None
        
        segments = parsed['segments']
        # Detected language
        detected_lang = parsed.get('language', lang or 'unknown')
        
        # Compute coverage stats
        hallucination_count = 0
        coverage = 0
        total_dur = max((s.get('end', 0) for s in segments), default=0)
        for s in segments:
            seg_dur = s.get('end', 0) - s.get('start', 0)
            word_count = len(s.get('words', []))
            if seg_dur > 15 and word_count < 5:
                hallucination_count += 1
            else:
                coverage += seg_dur
        coverage_pct = (coverage / total_dur * 100) if total_dur else 0
        
        total_words = sum(len(s.get('words', [])) for s in segments)
        print(f"   ✅ Gemini 3 Pro: {total_words} words → {len(segments)} segments, "
              f"lang={detected_lang}, coverage={coverage_pct:.0f}%", file=sys.stderr)
        
        return {
            "language": detected_lang,
            "language_probability": 0.95,
            "segments": segments,
            "_provider": "gemini-3-pro",
            "_hallucination_count": hallucination_count,
            "_coverage_pct": coverage_pct,
        }
    
    except urllib.error.HTTPError as e:
        err_body = e.read().decode()[:500] if hasattr(e, 'read') else ''
        print(f"   ❌ Gemini HTTP {e.code}: {err_body}", file=sys.stderr)
        return None
    except Exception as e:
        print(f"   ❌ Gemini fallback exception: {e}", file=sys.stderr)
        import traceback
        traceback.print_exc(file=sys.stderr)
        return None


def transcribe_with_soniox(audio_path, lang=None, filename_hint=None):
    """Soniox stt-async-v4 audio transcription — primary STT.
    
    Prednosti:
    - Najboljša natančnost za 60+ jezikov vključno s slovenščino
    - Brezhibno za narodno-zabavno glasbo (Avsenik, Modrijani, itd.)
    - Word-level timestamps + punctuation + diakritike
    - $0.10/h ($0.005 za 3-min pesem) - 4x cenejši kot Scribe
    - 4-13s za 180s audio (5x hitrejši kot Scribe)
    """
    import urllib.request
    import urllib.error
    
    api_key = os.environ.get("SONIOX_API_KEY")
    if not api_key:
        print(f"   ❌ SONIOX_API_KEY missing", file=sys.stderr)
        return None
    
    BASE = "https://api.soniox.com"
    print(f"🎤 Soniox stt-async-v4 transcribing {audio_path}...", file=sys.stderr)
    
    file_id = None
    trans_id = None
    
    def api_call(method, path, **kwargs):
        headers = kwargs.pop('headers', {})
        headers['Authorization'] = f'Bearer {api_key}'
        data = kwargs.get('data')
        if isinstance(data, dict):
            data = json.dumps(data).encode()
            headers['Content-Type'] = 'application/json'
        req = urllib.request.Request(f"{BASE}{path}", data=data, headers=headers, method=method)
        with urllib.request.urlopen(req, timeout=120) as resp:
            content = resp.read().decode()
            return json.loads(content) if content else {}
    
    try:
        # 1. Upload file (multipart)
        boundary = "----WebKitFormBoundary7MA4YWxkTrZu0gW"
        with open(audio_path, 'rb') as f:
            audio_bytes = f.read()
        body = b''.join([
            f"--{boundary}\r\n".encode(),
            b'Content-Disposition: form-data; name="file"; filename="audio.mp3"\r\n',
            b'Content-Type: audio/mpeg\r\n\r\n',
            audio_bytes,
            f"\r\n--{boundary}--\r\n".encode()
        ])
        req = urllib.request.Request(
            f"{BASE}/v1/files",
            data=body,
            headers={
                'Authorization': f'Bearer {api_key}',
                'Content-Type': f'multipart/form-data; boundary={boundary}',
            },
            method='POST'
        )
        with urllib.request.urlopen(req, timeout=120) as resp:
            file_data = json.loads(resp.read().decode())
        file_id = file_data['id']
        size_mb = len(audio_bytes) / 1024 / 1024
        print(f"   ✓ Uploaded {size_mb:.1f}MB → file_id={file_id}", file=sys.stderr)
        
        # 2. Create transcription
        config = {
            "model": "stt-async-v4",
            "file_id": file_id,
            "enable_language_identification": True,
        }
        # Language hints — prepoznaj jezik iz filename ali parametra
        if lang:
            config["language_hints"] = [lang]
        else:
            # Auto-detect iz filename
            fn_lower = (filename_hint or "").lower()
            if any(k in fn_lower for k in ["ansambel", "avsenik", "fehtar", "modrijan", "polka", "valček", "slovensk"]):
                config["language_hints"] = ["sl"]
            else:
                # Multilingual default - top svetovni
                config["language_hints"] = ["en", "sl", "de", "hr", "es", "fr", "it"]
        
        trans_data = api_call("POST", "/v1/transcriptions", data=config)
        trans_id = trans_data['id']
        print(f"   ✓ Transcription started: {trans_id}", file=sys.stderr)
        
        # 3. Poll status
        t0 = time.time()
        while True:
            status_data = api_call("GET", f"/v1/transcriptions/{trans_id}")
            status = status_data.get('status', 'unknown')
            elapsed = time.time() - t0
            if status == "completed":
                print(f"   ✓ Completed in {elapsed:.0f}s", file=sys.stderr)
                break
            if status == "error":
                print(f"   ❌ Soniox error: {status_data.get('error_message', '?')}", file=sys.stderr)
                return None
            if elapsed > 180:
                print(f"   ⚠️ Timeout (180s)", file=sys.stderr)
                return None
            time.sleep(2)
        
        # 4. Get transcript
        transcript_data = api_call("GET", f"/v1/transcriptions/{trans_id}/transcript")
        
        # Convert Soniox format → naš standard format (segments + words)
        tokens = transcript_data.get('tokens', [])
        if not tokens:
            print(f"   ❌ Empty transcript", file=sys.stderr)
            return None
        
        # Group tokens into words (Soniox vrača subwords; "Del" + " neb" + "a" = "Del neba")
        # Soniox token ima text in start_ms/end_ms. Beseda začne kjer ima text začetni space ali je prvi.
        words = []
        current_word = None
        for tok in tokens:
            text = tok.get('text', '')
            start_s = tok.get('start_ms', 0) / 1000
            end_s = tok.get('end_ms', 0) / 1000
            # Token, ki začne z space ali je <end>/special, je nova beseda
            if text.startswith(' ') or text in ('<end>', '<fin>'):
                if current_word and current_word['text'].strip():
                    words.append(current_word)
                if text in ('<end>', '<fin>'):
                    current_word = None
                    continue
                current_word = {'text': text, 'start': start_s, 'end': end_s, 'language': tok.get('language', lang or 'sl')}
            else:
                if current_word is None:
                    current_word = {'text': text, 'start': start_s, 'end': end_s, 'language': tok.get('language', lang or 'sl')}
                else:
                    # Append k current_word
                    current_word['text'] += text
                    current_word['end'] = end_s
        if current_word and current_word['text'].strip():
            words.append(current_word)
        
        # Group words into segments (po pavzah > 0.6s)
        segments = []
        if words:
            current_seg = {'start': words[0]['start'], 'end': words[0]['end'],
                          'text': words[0]['text'].strip(),
                          'words': [{'start': words[0]['start'], 'end': words[0]['end'], 'text': words[0]['text'].strip()}]}
            
            for w in words[1:]:
                gap = w['start'] - current_seg['end']
                if gap > 0.6 and len(current_seg['words']) >= 3:
                    segments.append(current_seg)
                    current_seg = {'start': w['start'], 'end': w['end'],
                                  'text': w['text'].strip(),
                                  'words': [{'start': w['start'], 'end': w['end'], 'text': w['text'].strip()}]}
                else:
                    current_seg['end'] = w['end']
                    current_seg['text'] = (current_seg['text'] + ' ' + w['text'].strip()).strip()
                    current_seg['words'].append({'start': w['start'], 'end': w['end'], 'text': w['text'].strip()})
            segments.append(current_seg)
        
        # Detected language
        detected_lang = lang or 'sl'
        if tokens:
            # Get most common language from tokens
            lang_counts = {}
            for tok in tokens:
                tl = tok.get('language')
                if tl:
                    lang_counts[tl] = lang_counts.get(tl, 0) + 1
            if lang_counts:
                detected_lang = max(lang_counts, key=lang_counts.get)
        
        # Compute coverage stats (compatible z ostalimi providerji)
        total_dur = max((s['end'] for s in segments), default=0)
        coverage = sum(s['end'] - s['start'] for s in segments)
        coverage_pct = (coverage / total_dur * 100) if total_dur else 0
        
        total_words = sum(len(s.get('words', [])) for s in segments)
        full_text = transcript_data.get('text', '')
        print(f"   ✅ Soniox: {total_words} words → {len(segments)} segments, "
              f"lang={detected_lang}, coverage={coverage_pct:.0f}%", file=sys.stderr)
        print(f"   📝 First 200 chars: {full_text[:200]!r}", file=sys.stderr)
        
        return {
            "language": detected_lang,
            "language_probability": 0.95,
            "segments": segments,
            "_provider": "soniox",
            "_hallucination_count": 0,  # Soniox redko halucinarala
            "_coverage_pct": coverage_pct,
        }
    
    except urllib.error.HTTPError as e:
        err_body = e.read().decode()[:500] if hasattr(e, 'read') else ''
        print(f"   ❌ Soniox HTTP {e.code}: {err_body}", file=sys.stderr)
        return None
    except Exception as e:
        print(f"   ❌ Soniox exception: {e}", file=sys.stderr)
        import traceback
        traceback.print_exc(file=sys.stderr)
        return None
    finally:
        # Cleanup — pošlji DELETE ampak ne preverjaj response (Soniox returns empty body)
        for path in ([f"/v1/transcriptions/{trans_id}"] if trans_id else []) + ([f"/v1/files/{file_id}"] if file_id else []):
            try:
                req = urllib.request.Request(f"{BASE}{path}",
                    headers={'Authorization': f'Bearer {api_key}'}, method='DELETE')
                urllib.request.urlopen(req, timeout=10)
            except Exception:
                pass


def transcribe_full(audio_path, lang=None, model_size="small", provider="auto", filename_hint=None):
    """STT dispatcher — Soniox primary z fallback chain.
    
    provider:
      - "soniox"     → Soniox stt-async-v4 (najboljši, $0.10/h, 5-15s)
      - "elevenlabs" → ElevenLabs Scribe ($0.40/h, 8-15s)
      - "gemini"     → Gemini 3 Pro ($3-5/h, 100-200s, najbolj točen za music)
      - "local"      → faster-whisper na CPU
      - "auto"       → Soniox primary, Scribe fallback, Gemini fallback ob halucinaciji
    """
    has_soniox = bool(os.environ.get("SONIOX_API_KEY"))
    has_scribe = bool(os.environ.get("ELEVENLABS_API_KEY"))
    has_gemini = bool(os.environ.get("GEMINI_API_KEY"))
    
    # Resolve "auto" → "soniox" če key, sicer fallback chain
    if provider in ("auto", "hybrid"):
        if has_soniox:
            provider = "soniox_chain"  # Soniox primary + fallbacks
        elif has_scribe and has_gemini:
            provider = "hybrid"  # legacy hybrid
        elif has_scribe:
            provider = "elevenlabs"
        else:
            provider = "local"
    
    # ─── SONIOX CHAIN: Soniox primary, Scribe/Gemini fallback ───
    if provider == "soniox_chain":
        print(f"🎯 Provider chain: Soniox → Scribe → Gemini", file=sys.stderr)
        result = transcribe_with_soniox(audio_path, lang=lang, filename_hint=filename_hint)
        
        if result and result.get("segments"):
            cov = result.get("_coverage_pct", 100)
            hall = result.get("_hallucination_count", 0)
            if cov >= 50 and hall == 0:
                return result
            print(f"   ⚠️ Soniox sumljiv (coverage {cov:.0f}%, hall {hall}) — try fallback", file=sys.stderr)
        else:
            print(f"   ❌ Soniox failed → fallback", file=sys.stderr)
        
        # Fallback 1: Scribe
        if has_scribe:
            print(f"   🔄 Fallback to Scribe...", file=sys.stderr)
            result2 = transcribe_with_elevenlabs(audio_path, lang=lang, filename_hint=filename_hint)
            if result2 and result2.get("segments"):
                cov = result2.get("_coverage_pct", 100)
                hall = result2.get("_hallucination_count", 0)
                if cov >= 50 and hall == 0:
                    return result2
                # ohrani za primerjavo
                result = result2 if not result else result
        
        # Fallback 2: Gemini (samo če sve doslej slabe)
        if has_gemini:
            print(f"   🔄 Fallback to Gemini 3 Pro (last resort)...", file=sys.stderr)
            result3 = transcribe_with_gemini(audio_path, lang=lang, filename_hint=filename_hint)
            if result3 and result3.get("segments"):
                return result3
        
        # Vrni karkoli imamo
        return result or {"language": "unknown", "language_probability": 0.0, "segments": []}
    
    # ─── SONIOX ONLY ───
    if provider == "soniox":
        if not has_soniox:
            return {"language": "unknown", "language_probability": 0.0, "segments": []}
        result = transcribe_with_soniox(audio_path, lang=lang, filename_hint=filename_hint)
        return result or {"language": "unknown", "language_probability": 0.0, "segments": []}
    
    # ─── HYBRID (legacy): Scribe primary, Gemini fallback ───
    if provider == "hybrid":
        if not has_scribe:
            provider = "gemini"
        else:
            # Try Scribe first
            print(f"🎯 HYBRID mode: Scribe primary, Gemini fallback", file=sys.stderr)
            result = transcribe_with_elevenlabs(audio_path, lang=lang, filename_hint=filename_hint)
            
            if result and result.get("segments"):
                hall_count = result.get("_hallucination_count", 0)
                cov_pct = result.get("_coverage_pct", 100)
                
                # Quality gate: če je Scribe rezultat dober, vrni ga
                if hall_count == 0 and cov_pct >= 50:
                    print(f"   ✅ Scribe OK (coverage {cov_pct:.0f}%) — no fallback needed", 
                          file=sys.stderr)
                    return result
                
                # Halucinacija ali nizko pokritje → preizkusi Scribe še 1x preden gremo na Gemini
                print(f"   ⚠️ Scribe quality issues (coverage {cov_pct:.0f}%, "
                      f"{hall_count} halu) — RETRY Scribe...", file=sys.stderr)
                result2 = transcribe_with_elevenlabs(audio_path, lang=lang, filename_hint=filename_hint)
                if result2 and result2.get("segments"):
                    h2 = result2.get("_hallucination_count", 0)
                    c2 = result2.get("_coverage_pct", 100)
                    if h2 == 0 and c2 >= 50:
                        print(f"   ✅ Scribe retry uspel: coverage {cov_pct:.0f}% → {c2:.0f}%", 
                              file=sys.stderr)
                        return result2
                    # Še vedno slabo, ali je drugi tek boljši?
                    if h2 < hall_count or c2 > cov_pct:
                        result = result2
                        hall_count = h2
                        cov_pct = c2
                
                # Še vedno halucinacija → Gemini fallback
                if has_gemini:
                    print(f"   🔄 Scribe še vedno slab (coverage {cov_pct:.0f}%, "
                          f"{hall_count} halu) — switching na Gemini 3 Pro...", file=sys.stderr)
                    gemini_result = transcribe_with_gemini(audio_path, lang=lang, filename_hint=filename_hint)
                    if gemini_result and gemini_result.get("segments"):
                        g_cov = gemini_result.get("_coverage_pct", 100)
                        g_hall = gemini_result.get("_hallucination_count", 0)
                        # Vzemi tisto kar je boljše
                        if g_hall < hall_count or g_cov > cov_pct:
                            print(f"   ✅ Gemini boljši: coverage {cov_pct:.0f}% → {g_cov:.0f}%, "
                                  f"hallu {hall_count} → {g_hall}", file=sys.stderr)
                            return gemini_result
                        else:
                            print(f"   ⚠️ Gemini ni boljši, ohrani Scribe", file=sys.stderr)
                            return result
                else:
                    print(f"   ⚠️ Gemini fallback ni dosegljiv — vrnem Scribe rezultat", 
                          file=sys.stderr)
                
                return result
            else:
                # Scribe popolnoma failed → Gemini direktno
                if has_gemini:
                    print(f"   🔄 Scribe failed → Gemini 3 Pro", file=sys.stderr)
                    gemini_result = transcribe_with_gemini(audio_path, lang=lang, filename_hint=filename_hint)
                    if gemini_result and gemini_result.get("segments"):
                        return gemini_result
                # Brez fallback → empty
                return {"language": "unknown", "language_probability": 0.0, "segments": []}
    
    # ─── GEMINI ONLY ───
    if provider == "gemini":
        if not has_gemini:
            print(f"   ❌ provider=gemini ampak GEMINI_API_KEY missing", file=sys.stderr)
            return {"language": "unknown", "language_probability": 0.0, "segments": []}
        result = transcribe_with_gemini(audio_path, lang=lang, filename_hint=filename_hint)
        if result and result.get("segments"):
            return result
        return {"language": "unknown", "language_probability": 0.0, "segments": []}
    
    # ─── ELEVENLABS / SCRIBE ONLY (z auto-retry) ───
    if provider == "elevenlabs" and has_scribe:
        result = transcribe_with_elevenlabs(audio_path, lang=lang, filename_hint=filename_hint)
        
        if result and result.get("segments"):
            hall_count = result.get("_hallucination_count", 0)
            cov_pct = result.get("_coverage_pct", 100)
            if hall_count > 0 or cov_pct < 50:
                print(f"   🔄 Halucinacija/nizko pokritje ({cov_pct:.0f}%, "
                      f"{hall_count} hallucination segs) — RETRY Scribe...", file=sys.stderr)
                result2 = transcribe_with_elevenlabs(audio_path, lang=lang, filename_hint=filename_hint)
                if result2 and result2.get("segments"):
                    h2 = result2.get("_hallucination_count", 0)
                    c2 = result2.get("_coverage_pct", 100)
                    if h2 < hall_count or c2 > cov_pct:
                        print(f"   ✅ Retry boljši: pokritje {cov_pct:.0f}% → {c2:.0f}%", 
                              file=sys.stderr)
                        result = result2
            return result
        return {"language": "unknown", "language_probability": 0.0, "segments": []}
    
    # ─── LOCAL faster-whisper ───
    return _transcribe_full_local(audio_path, lang=lang, model_size=model_size)


def _transcribe_full_local(audio_path, lang=None, model_size="small"):
    """Whisper transcript celega avdia. lang=None → robust auto-detect.
    
    Vrne empty transcript če Whisper ne najde govora (popolnoma instrumental)."""
    from faster_whisper import WhisperModel

    print(f"🧠 Whisper LOCAL {model_size}, lang={lang or 'auto'}", file=sys.stderr)
    m = WhisperModel(model_size, device="cpu", compute_type="int8")

    # Auto-detect z 3-sample voting da se zaklenemo na en jezik
    if not lang:
        print("   🔍 Robust lang detection (3 samples)...", file=sys.stderr)
        try:
            duration_proc = subprocess.run(
                ["ffprobe", "-v", "error", "-show_entries", "format=duration",
                 "-of", "default=nw=1:nokey=1", audio_path],
                capture_output=True, text=True
            )
            audio_duration = float(duration_proc.stdout.strip())
        except Exception:
            audio_duration = 180.0

        lang_votes = {}
        for ss in [max(15, audio_duration * 0.15), audio_duration * 0.45, audio_duration * 0.75]:
            if ss + 5 > audio_duration:
                continue
            sample = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
            sample.close()
            try:
                subprocess.run(
                    ["ffmpeg", "-y", "-ss", str(ss), "-i", audio_path,
                     "-t", "30", "-vn", "-ac", "1", "-ar", "16000",
                     "-c:a", "pcm_s16le", sample.name],
                    check=True, capture_output=True
                )
                _, sample_info = m.transcribe(sample.name, language=None, vad_filter=False)
                sl, sp = sample_info.language, float(sample_info.language_probability)
                lang_votes[sl] = lang_votes.get(sl, 0) + sp
                print(f"      sample @ {ss:.0f}s: {sl} (p={sp:.2f})", file=sys.stderr)
            except Exception as e:
                print(f"      sample @ {ss:.0f}s: failed", file=sys.stderr)
            finally:
                try:
                    os.unlink(sample.name)
                except Exception:
                    pass

        if lang_votes:
            lang = max(lang_votes.items(), key=lambda x: x[1])[0]
            print(f"   ✅ Lang lock: {lang}", file=sys.stderr)

    try:
        segs, info = m.transcribe(
            audio_path,
            language=lang,
            word_timestamps=True,
            # VAD filter kdaj izpusti vokal med glasbo — pri pesmi bolje brez
            vad_filter=False,
            # Anti-halucinacije
            condition_on_previous_text=False,
            temperature=0.0,
            compression_ratio_threshold=2.4,
            log_prob_threshold=-1.0,
            no_speech_threshold=0.6,
            # Beam search namesto greedy = bolj zanesljiv decode (manj halucinacij)
            beam_size=5,
            # Halucinacija detection: če je tišina dolga, ne pretvarjaj v tekst
            hallucination_silence_threshold=2.0,
        )
        detected_lang = info.language
        detected_prob = float(info.language_probability)
    except (ValueError, RuntimeError) as e:
        # Whisper failure (např. pri popolnoma instrumentalnih datotekah z VAD)
        print(f"   ⚠️ Whisper transcribe failed: {e}", file=sys.stderr)
        return {
            "language": "unknown",
            "language_probability": 0.0,
            "segments": [],
        }

    print(f"   Detekcija: {detected_lang} (p={detected_prob:.2f})", file=sys.stderr)

    segments = []
    for s in segs:
        words = []
        if s.words:
            for w in s.words:
                words.append({
                    "start": w.start,
                    "end": w.end,
                    "text": w.word,
                })
        segments.append({
            "start": s.start,
            "end": s.end,
            "text": s.text.strip(),
            "words": words,
        })

    return {
        "language": detected_lang,
        "language_probability": detected_prob,
        "segments": segments,
    }


def compute_energy_profile(audio_path, window_sec=1.0):
    """RMS dB na window_sec sekund. Vrne list (timestamp, rms_db)."""
    cmd = [
        "ffmpeg", "-i", audio_path,
        "-af", f"asetnsamples=n={int(16000 * window_sec)}:p=0,"
               f"astats=metadata=1:reset={window_sec},"
               f"ametadata=print:key=lavfi.astats.Overall.RMS_level:file=-",
        "-f", "null", "-",
    ]
    result = subprocess.run(cmd, capture_output=True, text=True)
    output = result.stdout + "\n" + result.stderr

    energies = []
    current_pts = 0.0
    for line in output.split("\n"):
        line = line.strip()
        m = re.search(r"pts_time:(\S+)", line)
        if m:
            try:
                current_pts = float(m.group(1))
            except ValueError:
                pass
            continue
        if "RMS_level=" in line:
            val = line.split("RMS_level=")[-1].strip()
            try:
                rms = float(val)
                # -inf zamenjamo z -90
                if rms < -90 or rms != rms:  # NaN check
                    rms = -90.0
                energies.append((current_pts, rms))
                current_pts += window_sec
            except ValueError:
                pass

    return energies


def detect_vocal_sections(segments, max_gap=3.0):
    """Združi consecutive segmente v "vokalne sekcije"."""
    if not segments:
        return []
    sections = []
    current = {
        "start": segments[0]["start"],
        "end": segments[0]["end"],
        "segments": [segments[0]],
        "text": segments[0]["text"],
    }
    for seg in segments[1:]:
        if seg["start"] - current["end"] > max_gap:
            sections.append(current)
            current = {
                "start": seg["start"],
                "end": seg["end"],
                "segments": [seg],
                "text": seg["text"],
            }
        else:
            current["end"] = seg["end"]
            current["segments"].append(seg)
            current["text"] += " " + seg["text"]
    sections.append(current)
    return sections


def avg_energy_in_range(energies, start, end):
    """Povprečna RMS v rangeu."""
    vals = [r for (t, r) in energies if start <= t <= end]
    if not vals:
        return -90.0
    return sum(vals) / len(vals)


def score_section_as_chorus(section, all_sections, energies, avg_rms):
    """Score sekcijo kot kandidat za refren.

    Faktorji:
    - Ponavljajoče besede (low unique-word-ratio) = refren
    - Visoka energija
    - Sekcija se pojavi večkrat v pesmi (refren se ponovi)
    - Krajše vrstice (3-8 besed)
    """
    text = section["text"].lower()
    words = re.findall(r"\b\w+\b", text)
    if not words:
        return 0

    unique_ratio = len(set(words)) / len(words)
    # Refren = nizko unique ratio (ponovitve)
    chorus_signal = max(0, (1.0 - unique_ratio) * 30)

    # Energija
    sec_energy = avg_energy_in_range(energies, section["start"], section["end"])
    energy_above = max(0, sec_energy - avg_rms)
    energy_score = energy_above * 8

    # Kako pogosto se pojavi podobno besedilo
    repeat_count = 0
    for other in all_sections:
        if other is section:
            continue
        other_text = other["text"].lower()
        other_words = set(re.findall(r"\b\w+\b", other_text))
        common = set(words) & other_words
        # Če imata >50% besed skupnih, je verjetno isti refren
        if len(common) >= len(set(words)) * 0.5 and len(common) >= 3:
            repeat_count += 1
    repeat_score = repeat_count * 25

    # Dolžina vrstice
    duration = section["end"] - section["start"]
    if 3 <= duration <= 25:
        length_score = 10
    elif duration > 25:
        length_score = 5
    else:
        length_score = 2

    return chorus_signal + energy_score + repeat_score + length_score


def find_chorus(transcript, energies, video_duration):
    """Najde najbolj verjeten refren."""
    sections = detect_vocal_sections(transcript["segments"])
    if not sections:
        return None

    avg_rms = sum(r for (_, r) in energies) / len(energies) if energies else -30.0

    candidates = []
    for sec in sections:
        score = score_section_as_chorus(sec, sections, energies, avg_rms)
        candidates.append({
            "start": sec["start"],
            "end": sec["end"],
            "duration": sec["end"] - sec["start"],
            "text_preview": sec["text"][:80],
            "score": round(score, 2),
            "avg_rms": round(avg_energy_in_range(energies, sec["start"], sec["end"]), 2),
        })

    # Sort by score descending
    candidates.sort(key=lambda c: -c["score"])

    if not candidates:
        return None

    return {
        "best": candidates[0],
        "all_candidates": candidates[:10],
        "avg_rms_total": round(avg_rms, 2),
    }


def smart_clip_range(chorus, transcript, video_duration,
                      target_duration=30, max_duration=45, min_duration=20,
                      include_prebuild=False):
    """Vrne clip range TOČNO kot je odločil LLM.
    
    Sistem NE razširja in NE skrajšuje LLM-ove odločitve. LLM ima ves
    transkript + lyrics z web_search in lahko razmišlja vsebinsko o:
    - kateri refren
    - koliko ponovitev (1, 2, 3?)
    - ali vključiti intro klic ('Ajmo Janezi!')
    - kje naravno konča
    
    Edina varnost: če Claude vrne nič, fallback na sredino videa.
    """
    if not chorus or not chorus.get("best"):
        # Fallback: sredina videa
        mid = video_duration / 2
        start = max(0, mid - target_duration / 2)
        return {
            "start": start,
            "end": min(video_duration, start + target_duration),
            "reason": "fallback_middle",
        }

    best = chorus["best"]
    actual_start = best["start"]
    actual_end = best["end"]
    
    # Samo varnostni cap: NIKOLI ne čez video duration
    actual_start = max(0, actual_start)
    actual_end = min(video_duration, actual_end)
    
    # Če je nekako reverse (start > end), popravi
    if actual_start >= actual_end:
        actual_end = min(video_duration, actual_start + target_duration)

    return {
        "start": round(actual_start, 2),
        "end": round(actual_end, 2),
        "duration": round(actual_end - actual_start, 2),
        "reason": "smart_chorus_with_prebuild" if include_prebuild else "smart_chorus_only",
        "chorus_start": round(best["start"], 2),
        "chorus_end": round(best["end"], 2),
    }


def detect_audio_fade(clip_range, transcript, video_duration=None):
    """Določi fade-in/fade-out trajanje + ev. razširi clip range, da fade
    ne reže besedila na koncu refrena.

    Logika:
    - Če clip začne sredi vokala → 0.5s fade in
    - Če se konča sredi vokala → razširi clip do konca segmenta (+ buffer),
      potem 1.0s fade out
    - Sicer manj fade
    """
    cs, ce = clip_range["start"], clip_range["end"]

    # Najdi segment, ki konča znotraj clip-a (ali je clip end znotraj segmenta)
    starts_in_vocal = False
    ends_in_vocal = False
    end_segment = None
    for seg in transcript["segments"]:
        if seg["start"] <= cs <= seg["end"]:
            starts_in_vocal = True
        if seg["start"] <= ce <= seg["end"]:
            ends_in_vocal = True
            end_segment = seg

    # Če clip konča znotraj segmenta, razširi do konca segmenta + 0.5s buffer
    extended_end = ce
    if end_segment:
        extended_end = end_segment["end"] + 0.5
        if video_duration is not None:
            extended_end = min(extended_end, video_duration)

    # Fade-in: če clip začne MED vokalom, fade-in mora biti zelo kratek
    # da ne odreže prve besede. Pri vokalnem začetku samo 0.05s "smooth click prevention",
    # ne pravi audible fade. Pri instrumentalnem intro lahko 0.2-0.3s.
    fade_in = 0.05 if starts_in_vocal else 0.2
    # Krajši fade out (0.5s) ker zdaj clip konča po koncu vokala
    fade_out = 0.3 if ends_in_vocal else 0.4

    return {
        "fade_in": fade_in,
        "fade_out": fade_out,
        "extended_end": round(extended_end, 2),
        "ends_in_vocal": ends_in_vocal,
    }


def _build_analysis_prompt(transcript, video_duration, target_duration=30, filename_hint=None, include_prebuild=False):
    """Pripravi enotni prompt za Claude/Gemini analizo.
    
    include_prebuild: če True, lahko vključi pre-chorus pred refrenom.
                      če False (default), MORA biti SAMO refren — strogo.
    """
    lines = []
    for seg in transcript["segments"]:
        start = seg["start"]
        end = seg["end"]
        text = seg["text"].strip()
        lines.append(f"[{start:6.1f}-{end:6.1f}] {text}")
    transcript_text = "\n".join(lines)

    hint_block = ""
    if filename_hint:
        hint_block = f"""

🎵 IME DATOTEKE: "{filename_hint}"

🚨 **PRVI KORAK — VEDNO PRED ANALIZO**:
Iz imena datoteke prepoznaj izvajalca + naslov pesmi. Potem **OBVEZNO uporabi web_search tool** da poiščeš pravo besedilo pesmi — TUDI ČE MISLIŠ DA POZNAŠ PESEM.

Razlog: večinoma ne poznaš celotnih besedil pesmi (predvsem ne-angleških). Brez pravega besedila NE MOREŠ:
- Pravilno prepoznati strukture (verz / pre-chorus / chorus / bridge)
- Vedeti kje refren **začne in konča** (vključno z outro frazami)
- Popraviti STT halucinacij

📋 **Search strategija** (univerzalna za vse jezike):
1. Prvo iskanje: `[izvajalec] [naslov] lyrics` ALI `[izvajalec] [naslov] besedilo/Songtext/letra/versuri`
2. Če ni rezultatov: `[del transkripta - 4-5 zaporednih besed] lyrics`
3. Trusted lyrics sajti po jezikih:
   - 🇸🇮 SLO: besedila.com, lyricstranslate.com
   - 🇩🇪 DE: songtexte.com, lyricstranslate.com
   - 🇭🇷🇷🇸 HR/SR/BS: tekstovi.net, lyricstranslate.com
   - 🇪🇸 ES: letras.com, musica.com
   - 🇷🇴 RO: versuri.ro, lyricstranslate.com
   - 🇮🇹 IT: angolotesti.it
   - 🇫🇷 FR: paroles.net
   - 🇬🇧🇺🇸 EN: genius.com, azlyrics.com
   - **Univerzalno**: lyricstranslate.com (vsi jeziki)

Ko najdeš lyrics:
- Identificiraj kateri del je REFREN (ponavlja se)
- Identificiraj VERZE (zgodba, ne ponavlja se)
- Identificiraj BRIDGE / PRE-CHORUS / OUTRO če obstajajo
- Mapiraj transkript timestamp-e na strukturne dele
- Popravi corrected_segments z dejanskim besedilom

🎯 **NASLOV PESMI — KAKO POVEZAN Z REFRENOM**:
Naslov pesmi je **včasih** v refrenu, **včasih** v verzu. NE silujte!
- POGOSTO v refrenu: "Pijan" → "pijan, pijan"; "Žena me tepe" → "Žena me tepe"; "Brajde" → "v brajde"
- VČASIH samo v verzu: "Cvetele so maline" → naslov je v VERZU, refren je drug ("Naj veter zdaj ponese...")
- **NE iskaj naslova v refrenu na silo** — refren je tisti del ki se **PONAVLJA** (2-3x), ne tisti ki vsebuje naslov

🎯 **KAKO RES IDENTIFICIRATI REFREN**:
1. Refren je tisti del besedila, ki se v pesmi **DEJANSKO PONOVI 2-3x z ENAKIM besedilom**
2. Verzi imajo **različno besedilo** vsakič (pripovedujejo zgodbo)
3. Najprej najdi besedilo, ki se ponavlja — TO je refren
4. Ko najdeš PRVI nastop ponavljajočega dela → tam začni clip

⚠️ **PAZI**: prvi verz pesmi se pogosto začne **takoj po intro-u** (5-15s) in je kontekstualen — TO NI REFREN. Refren običajno pride **po prvem verzu** (pri 30-60s, odvisno od pesmi).
"""

    return f"""Tu je transcript pesmi iz STT modela (timestamp v sekundah, besedilo):

{transcript_text}

Cela pesem traja {video_duration:.1f}s. Cilj: izrezati ~{target_duration}s odsek za TikTok/Instagram Reel.{hint_block}

⚠️ POMEMBNO: STT lahko naredi napake v vseh jezikih, posebej:
- Pri narečjih, slovanskih jezikih, romanskih jezikih
- Generira "tipičen" tekst (npr. tekst druge pesmi istega izvajalca)
- Lahko vstavi besede ki se POdoBNO slišijo, ampak imajo ČISTO drug pomen

KAKO PREPOZNATI HALUCINACIJO:
- Tekst nima smisla v kontekstu pesmi
- Različni segmenti imajo nepovezane teme (kot da bi bilo več pesmi)
- Refren je v vsakem ponovitvi različen (refren se MORA ponavljati identično)
- Tekst je premalo **glede na trajanje** (več tišine = manj besed, ne več)

PROSIM:
1. Preberi celoten tekst in razumi strukturo (intro / verz / pre-chorus / refren / bridge / outro)
2. POPRAVI očitne halucinacije:
   - Če prepoznaš pesem (po izvajalcu, naslovu, znaku besedila) → **uporabi PRAVO besedilo**
   - Če halucinacijo ne moreš popraviti, **odstrani segment** (raje brez podnapisa kot napačen)
   - Refren MORA imeti vse pojavitve ENAKE
   - Popravi pomešane jezike (vse vrstice v enem jeziku)
   - Ohrani timestamp-e nespremenjene
3. Prepoznaj REFREN: del besedila ki se PONAVLJA (ponavadi 2-4 vrstice, ki se v pesmi večkrat ponovijo). To je **univerzalno za vse jezike** — refren je strukturni element pesmi, ne le slovenske/nemške/angleške.

{"" if include_prebuild else '''4. **🎯 IZBIRA REFRENA — VSEBINSKO RAZMIŠLJANJE**

   TI ODLOČAŠ na osnovi vsebine, ritma, energije pesmi.
   Sistem ne bo razširil ne skrajšal tvoje izbire — kar vrneš, to se uporabi.
   
   ## CILJ: ~30 sekund (tipično za TikTok/Instagram Reel)
   
   Dolžina je SVOBODNA glede na strukturo pesmi:
   - 12-15s = en kratki refren (če je pesem zelo kratka)
   - **20-35s = dva zaporedna refrena (NAJBOLJŠA opcija!)** — največkrat idealno
   - 30-40s = refren + drug refren če sta vsebinsko povezana
   
   ## 🎨 SEBASTJANOV FILING (kuratorski občutek za TikTok/Reels):
   
   Sebastjan je TV producent ki dela reele za FOLX TV. Ima razvit občutek **kaj gre v reel** in **kaj ne**.
   Tu je njegov filing strnjen v pravila:
   
   ### ✅ DA vključi:
   - **Intro klici/fraze** ki vodijo v refren: "Ajmo Janezi!", "Ajmo Slovenci!", "Hey!", "Yeah!", "Pa-pa!"
   - **Dva zaporedna refrena** = najboljša opcija (~30s) — to je "filing 1.0"
   - **Naravni outro filler** na koncu refrena (kratko: "yeah", "aj-aj", "uh la la") — če **NI predolg**
   - **Pevčev držan ton** 1-3s — to je del refrena, naravni dah pesmi
   
   ### ❌ NE vključi:
   - **Verzov/kitic** — pripovedujejo zgodbo, niso catchy za reel
   - **Pre-chorus** build-up pred refrenom (razen če uporabnik eksplicitno izbere)
   - **2 različnih refrenov pomešanih** (če pesem ima refren A in refren B, vzemi samo enega)
   - **Dolg outro filler** v fade-out (5+ sekund "o o o", "la la la") — to je konec pesmi, NE refren
   - **Instrumental break** — ni vokalov, ni vsebina
   
   ### 🎯 Sebastjanovo razumevanje refrena:
   - **Refren = del besedila ki se v pesmi PONOVI 2-3x z ENAKIM besedilom**
   - **NI nujno** da refren vsebuje naslov pesmi (npr. "Cvetele so maline" — naslov je v VERZU, refren je drug)
   - **Pri hitrih pop pesmih**: refren je catchy "hook" del z melodijo
   - **Pri počasnih NZ**: refren je melodičen del po verzu (Avsenik tip)
   - **VEDNO uporabi PRVI nastop refrena**, ne 2. ali 3.
   
   ### 📏 Občutek za dolžino:
   - **Hitri komadi** (BRAJDE, PA PA, PIJAN, FICKA): 1-2 refrena ~25-30s
   - **Počasne NZ** (CVETELE, GORENJSKA, DOMOTOŽJE): tipično 1 polni refren ~15-20s
   - **Schlager/pop** (Abracadabra, Žena me tepe): 2 zaporedna refrena ~30s
   - **NE razmišljaj o sekundah** — razmišljaj o **VSEBINI**
   
   ## 🎯 KRITIČNO: clip start = TOČNO prva beseda PRVE vrstice refrena
   - **OPOZORILO**: Soniox lahko združi konec verza z začetkom refrena v en segment (npr. "[43-47s] doma. V Ljubljani se"). Uporabi WORD-LEVEL timestampe da najdeš točno **prvo besedo refrena**!
   - **NIKOLI** ne začni sredi besede ali sredi vrstice refrena
   
   ## 🎯 KRITIČNO: clip end = po izpevu zadnje besede refrena
   - **OPOZORILO**: Soniox lahko zadnji izpeti ton ("oba", "doma", "srca") razpiše kot 15-20s segment zaradi instrumental fade-out. **TO NI BESEDILO** — odreži 1-2s po zadnji besedi
   - **NE** vključuj dolgega outro filler-ja ("o o o", "la la la" 5+s)
   
   ## 📚 SEBASTJANOVI 15 PRIMEROV — uporabi kot REFERENCO:
   
   | Pesem | Refren (kar ide v reel) | Občutek |
   |---|---|---|
   | **BRAJDE** (FIRBCI) | "Ajmo Janezi! Pejd' greva, ajde, na traktorju od Majde, noben naju ne najde, ko peljem te v brajde. Da v senci hladni vžgem te po tazadn'i, ko s tebe trgam cunje, Gucci, Louis Vuitton!" | 2x ponovitev z "Ajmo Janezi" intro = 28s |
   | **CVETELE SO MALINE** (Avsenik) | "Naj veter zdaj ponese moje sanje čez del neba, tja, kjer je ona doma. Zašepetaj ji bog tja pod kostanje, da sanjam še, kar sva nekoč oba." | PRVI nastop, ~15s, **VKLJUČI vse do konca "oba" (zadnja beseda IZPETO)** — če je instrumental break za besedo "oba", clip še VEDNO mora vključiti TO BESEDO + 0.5s po njej (NE odreži pred "oba"!) |
   | **ENA BOLHA ZA POMOČ** (Avsenik) | "Kam se ti tako mudi, punca, voda ne gori, daj počakaj še malo! Mlada ledi, k' si in frej, dej se z fanti mal' igrej, živeti je treba zdej-ej-ej. Hvala, lupčka, lahko noč, ena bolha za pomoč in užiVAJ življenje. Dobro jutro, dober dan, saj ljubezen pač ni plan, in užiVAJ samski stan." | Dolg refren, ~30s |
   | **ŽENA ME TEPE** (Avsenik) | "Žena me tepe, mi prazni žepe, da vidi, kje in s kom sem bil. Jaz pa razlagam, da ne vem, da nisem kriv, da šel naravnost sem domov, pa se izgubil. Žena me tepe, obrača žepe, preverja, kje in s kom sem pil. Saj bi ji rekel, naj se že enkrat ustav', pa vem, da ima po svoje prav." | 2 vrstici refrena ~30s |
   | **STISN SE K MEN** (Stil) | "Stisni se k meni, tak ni noben. Ko bova stara, bova spala, do takrat pa boogie-woogie žgala, ne? Ne hecam se, ko se zaljubim, ljubim, kot da zadnjič je." | 2x ponovitev = ~30s |
   | **PA PA** (Unikat) | "Pa-pa! Sem rekla in sem mu pomahala. Bla-bla! Dovolj mi je na dušo pihanja. Pa-pa! Veliko takih sem že srečala. Ne, ne, o ne, takim ne dam srca." | 1 polni refren ~15s, 2x = ~30s |
   | **GOJZAR TANC** (Fehtarji) | "Hojladijo, v hribe gremo, gore kličejo, vrhovi vabijo. Hojladijo, v hribe gremo, v lepi gorski raj še pridemo nazaj." | 1-2 nastopov ~25s |
   | **GORENJSKA LJUBLJENA** (Fehtarji) | "Kjer gore se dotikajo neba. Tu, kjer sem rojen, je najlepši del sveta, doline izpod Triglava. Tu je moja mati nekoč bila doma, gorenjska ljubljena." | Refren začne **sredi vrstice** "...gorenjski raj, KJER gore..." — start je beseda "kjer" |
   | **PODEŽELSKI ROCK'N'ROLL** (Fehtarji) | "Na ta rdečmu Fergusonu sem jo gor naslonu, to je podeželski rock'n'roll. U kabini ga s' tiščala, ne da bi se not peštala, sam' All Star-ke sva še mela gor. Noge dviga prot' oblakom, nafta šla je žvižgat k rakom, ker sem pustu vžgan motor." | KONČA pri "vžgan motor", **NE vključi "O o o" outro** |
   | **DOMOTOŽJE V POMLADI** (Fantje z vseh vetrov) | "Oj deželica ti mila, moj najlepši beli cvet, ti ljubezen si vsadila mi v otroštvo mladih let. Bele breze pa šumijo, vabijo me le domov, meni se oči solzijo, ko želijo si domov." | Pre-chorus + refren skupaj ~30s |
   | **KO MISLIM NATE** (GADI) | "Ko mislim, mislim nate, vse na svetu dal bi zate, in le na tvoje ime mi noro tolče srce. Ko mislim, mislim nate, do sonca peš odšel bi zate, v mojih mislih bilo, ni še druge in nikdar je ne bo." | 2x "Ko mislim mislim nate" = ~30s |
   | **PIJAN** (GADI) | "Pijan, pijan, pijan od tebe in sreče. Pijan, pijan, s teboj do belega dne. Pijan, pijan, pijan od tebe in sreče. Pijan, pijan, s teboj do belega dne." | 2x ponovitev ~25s |
   | **ABRACADABRA** (Lady Gaga) | "Abracadabra, amor-ooh-na-na, abracadabra, morta-ooh-ga-ga. Abracadabra, abra-ooh-na-na, in her tongue she said, 'Death or love tonight.'" | 1-2 nastopov ~30s |
   | **FICKA** (Sarah Connor) | "Ist mir scheißegal, was ihr schreibt, ihr Ficka, denn irgendwer hat immer, immer, immer was zu meckern. An manchen Tagen geht's mir nah, und es ärgert mich, denn nichts und wieder nichts davon ist wahr. Ich mach mir die Welt wie 'ne Wiese, mir gefällt. Scheiß auf falsche Freunde und scheiß auf euer Geld. Willkommen, willkommen in meiner wunderschönen Welt." | Dolg refren ~35s |
   | **NA LEPO SKUPNO POT** (Vikend) | "Na lepo skupno pot sta stopila, združila prstana sta dvoje src. Ljubezen vaju bo povsod vodila, naj sonce sreče greje vajin vrt." | 1 polni refren ~15s, 2x = ~30s |
   
   **POMEMBNO**: ti primeri so Sebastjanov filing. Pri **NOVIH pesmih** (ki niso v tabeli) **POSNEMI ta filing**:
   - Najdi del besedila ki se ponovi 2-3x
   - Vključi naravne intro klice če obstajajo
   - NE vključuj verzov/kitic
   - NE vključuj dolgega outro filler-ja
   - NE mešaj 2 različnih refrenov
   - Cilj ~30s, fleksibilno glede na pesem
'''}{'''4. **IZBERI ODSEK — REFREN + PRE-CHORUS:**

   Uporabnik je izbral način "**REFREN + PRE-CHORUS**". 

   ## OBVEZNO: cel **PRVI** refren (kot opisano spodaj)
   
   ## OPCIJSKO: pre-chorus PRED refrenom
   - **Pre-chorus = zadnja 1-2 vrstici verza tik pred refrenom** (slišne, povezane z refrenom)
   - **Dodaj samo če**:
     - Je tik pred refrenom (brez pavze ali instrumental vmes)
     - Vsebinsko vodi v refren (gradnja občutka)
     - Je kratek: 4-10 sekund
   - **Ne dodajaj** če bi presegel skupno dolžino 35s
   
   ## REFREN — kot pri "samo refren":
   - Začetek refrena = prva vrstica refrena
   - Konec refrena = vključno z vsemi outro frazami in zadnjim držečim tonom
   - Naravni izpev (ej-ej-ej, oh oh, la la la, etc.)
   
   ## Skupna dolžina: 18-35 sekund
''' if include_prebuild else ""}

5. Če transkript je v večini halucinacija (manj kot 30% smiselnih besed), v "reason" napiši "STT_HALLUCINATION_DETECTED"

Odgovori SAMO v JSON formatu (brez markdown, brez razlage):
{{
  "start": <sekunde>,
  "end": <sekunde>,
  "reason": "<kratka razlaga>",
  "chorus_text": "<besedilo refrena>",
  "structure": "<1 stavek o strukturi pesmi>",
  "language": "<jezik: sl/de/hr/bs/sr/en/it/es/fr>",
  "hallucination_detected": <true/false>,
  "corrected_segments": [
    {{"start": <s>, "end": <s>, "text": "<popravljeno besedilo ALI prazno če halucinacija>"}}
  ]
}}

V "corrected_segments" vključi VSE segmente iz inputa s popravljenim besedilom. Halucinacije nadomesti s pravim besedilom (če veš) ALI pusti prazno besedilo."""


def _parse_llm_response(text, video_duration):
    """Parse JSON odgovor iz LLM-a, vrne None če invalid."""
    text = text.strip()
    # Odstrani markdown ovoj če obstaja
    if text.startswith("```"):
        text = re.sub(r"^```(?:json)?\s*", "", text)
        text = re.sub(r"\s*```$", "", text)
    # Včasih je pred JSON-om še kakšna razlaga, vzemi prvi { ... } blok
    first_brace = text.find("{")
    last_brace = text.rfind("}")
    if first_brace >= 0 and last_brace > first_brace:
        text = text[first_brace:last_brace + 1]

    result = json.loads(text)

    start = float(result["start"])
    end = float(result["end"])
    if start >= end or start < 0 or end > video_duration:
        print(f"   ⚠️ LLM returned invalid range: {start}-{end}", file=sys.stderr)
        return None

    return {
        "start": round(start, 2),
        "end": round(end, 2),
        "duration": round(end - start, 2),
        "reason": result.get("reason", ""),
        "chorus_text": result.get("chorus_text", ""),
        "structure": result.get("structure", ""),
        "language": result.get("language"),
        "corrected_segments": result.get("corrected_segments"),
    }


def analyze_with_claude(transcript, video_duration, target_duration=30, model="claude-sonnet-4-6", filename_hint=None, include_prebuild=False):
    """Pošlje transkript Claude API-ju (Anthropic).
    
    model: claude-sonnet-4-6 (default), claude-haiku-4-5-20251001, claude-opus-4-7
    filename_hint: ime datoteke (Claude lahko prepozna pesem in popravi halucinacije)
    """
    api_key = os.environ.get("ANTHROPIC_API_KEY")
    if not api_key:
        print("   ⚠️ ANTHROPIC_API_KEY ni nastavljen — preskakujem Claude analizo", file=sys.stderr)
        return None

    if not transcript.get("segments"):
        return None

    prompt = _build_analysis_prompt(transcript, video_duration, target_duration, filename_hint=filename_hint, include_prebuild=include_prebuild)

    try:
        import urllib.request
        import urllib.error

        # Initial messages
        messages = [{"role": "user", "content": prompt}]

        # Sonnet 4.6 podpira web_search tool — Claude lahko poišče prave lyrics
        # za pesmi v slovenščini/hrvaščini/itd., če jih ne pozna iz training data.
        tools = [{
            "type": "web_search_20250305",
            "name": "web_search",
            "max_uses": 3,  # Maksimalno 3 search-i = $0.03/job
        }]

        # Agentic loop: Claude lahko kliče web_search, dobi rezultate, vrne final answer
        max_iterations = 5
        for iteration in range(max_iterations):
            body = json.dumps({
                "model": model,
                "max_tokens": 8192,
                "messages": messages,
                "tools": tools,
            }).encode("utf-8")

            req = urllib.request.Request(
                "https://api.anthropic.com/v1/messages",
                data=body,
                headers={
                    "Content-Type": "application/json",
                    "x-api-key": api_key,
                    "anthropic-version": "2023-06-01",
                },
                method="POST",
            )
            with urllib.request.urlopen(req, timeout=180) as resp:
                data = json.loads(resp.read().decode("utf-8"))

            content = data.get("content", [])
            if not content:
                print("   ⚠️ Claude vrnil prazen odgovor", file=sys.stderr)
                return None

            stop_reason = data.get("stop_reason")
            if stop_reason == "max_tokens":
                usage = data.get("usage", {})
                print(
                    f"   ⚠️ Claude odrezan (max_tokens): "
                    f"input={usage.get('input_tokens')} output={usage.get('output_tokens')}",
                    file=sys.stderr,
                )
                return None

            # Če je end_turn → smo končali, parsiraj text
            if stop_reason in ("end_turn", "stop_sequence"):
                # Najdem zadnji text block
                text_blocks = [b for b in content if b.get("type") == "text"]
                if text_blocks:
                    text = text_blocks[-1].get("text", "").strip()
                    break
                print("   ⚠️ Claude end_turn brez text bloka", file=sys.stderr)
                return None

            # Če je tool_use → Claude kliče web_search; appendamo response in nadaljujemo
            if stop_reason == "tool_use":
                # Anthropic web_search tool je server-side — sami obdela searches in vrne web_search_tool_result
                # Ampak v API odgovoru so OBA: tool_use IN web_search_tool_result kot del content
                # Torej končni text že obstaja v naslednji iteraciji
                # Appendamo content do messages in pošljem nazaj (Claude bo nadaljeval)
                messages.append({"role": "assistant", "content": content})
                # Claude server-side že obdela search, samo nadaljujemo s pustim user msg
                # Ampak server-side tools NE potrebujejo follow-up tool_result
                # Pravilen flow: če stop_reason=tool_use ampak web_search_tool_result je že v content,
                # potem Claude sam nadaljuje. Drugače moramo poslati tool_result.
                
                # Preverim ali so že rezultati v content
                has_results = any(b.get("type") == "web_search_tool_result" for b in content)
                if has_results:
                    # Server-side: Anthropic je sam obdelal search, čakamo nadaljevanje
                    # Pošlji nazaj brez sprememb da Claude nadaljuje
                    print(f"   🔍 Claude je iskal lyrics, čakam nadaljevanje (iter {iteration+1})", file=sys.stderr)
                    continue
                else:
                    print(f"   ⚠️ tool_use brez results", file=sys.stderr)
                    return None
            
            # Drugi stop reasons
            print(f"   ⚠️ Nepričakovan stop_reason: {stop_reason}", file=sys.stderr)
            return None
        else:
            print(f"   ⚠️ Presežena max_iterations ({max_iterations})", file=sys.stderr)
            return None

        result = _parse_llm_response(text, video_duration)
        if not result:
            return None

        print(f"   🤖 Claude ({model}) izbral: {result['start']:.1f}-{result['end']:.1f}s", file=sys.stderr)
        print(f"      Razlog: {result.get('reason', '')[:80]}", file=sys.stderr)
        print(f"      Struktura: {result.get('structure', '')[:80]}", file=sys.stderr)
        if result.get("corrected_segments"):
            print(f"      Popravljeni segmenti: {len(result['corrected_segments'])}", file=sys.stderr)

        result["source"] = f"claude:{model}"
        return result
    except urllib.error.HTTPError as e:
        body = e.read().decode("utf-8", errors="replace")[:500]
        print(f"   ❌ Claude API HTTP {e.code}: {body}", file=sys.stderr)
        return None
    except Exception as e:
        print(f"   ❌ Claude analysis failed: {e}", file=sys.stderr)
        return None


def analyze_with_gemini(transcript, video_duration, target_duration=30, model="gemini-3.1-pro-preview", filename_hint=None, include_prebuild=False):
    """Pošlje transkript Gemini API-ju (Google).
    
    Gemini 3.1 Pro ima najboljši multilingual rezultat (MMMLU 92.6%) — odličen za SLO/HR/BS.
    """
    api_key = os.environ.get("GEMINI_API_KEY") or os.environ.get("GOOGLE_API_KEY")
    if not api_key:
        print("   ⚠️ GEMINI_API_KEY ni nastavljen — preskakujem Gemini analizo", file=sys.stderr)
        return None

    if not transcript.get("segments"):
        return None

    prompt = _build_analysis_prompt(transcript, video_duration, target_duration, filename_hint=filename_hint, include_prebuild=include_prebuild)

    try:
        import urllib.request
        import urllib.error

        url = f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={api_key}"
        # Gemini 3.x Pro je THINKING model — porabi tokene tudi za internal reasoning (thoughtsTokenCount).
        # 4096 je prenizko: pri velikih transkriptih thinking lahko porabi 1500-3000 tokenov,
        # output (corrected_segments za 60+ segmentov) pa še dodatnih 3000-7000 → odreže JSON na pol
        # (finishReason: MAX_TOKENS) in vrne nepopolen, neveljaven JSON.
        # 32768 daje dovolj prostora za thinking + cel JSON output tudi pri dolgih pesmih.
        body = json.dumps({
            "contents": [{
                "role": "user",
                "parts": [{"text": prompt}],
            }],
            "generationConfig": {
                "temperature": 0.1,
                "maxOutputTokens": 32768,
                "responseMimeType": "application/json",
            },
        }).encode("utf-8")

        req = urllib.request.Request(
            url,
            data=body,
            headers={"Content-Type": "application/json"},
            method="POST",
        )
        with urllib.request.urlopen(req, timeout=180) as resp:
            data = json.loads(resp.read().decode("utf-8"))

        candidates = data.get("candidates", [])
        if not candidates:
            print("   ⚠️ Gemini vrnil 0 candidates", file=sys.stderr)
            return None

        cand0 = candidates[0]
        finish_reason = cand0.get("finishReason", "?")
        usage = data.get("usageMetadata", {})

        # Diagnostika: če je finishReason == MAX_TOKENS, je output odrezan in JSON je invalid
        if finish_reason == "MAX_TOKENS":
            print(
                f"   ⚠️ Gemini odrezan (MAX_TOKENS): "
                f"prompt={usage.get('promptTokenCount')} "
                f"thoughts={usage.get('thoughtsTokenCount')} "
                f"output={usage.get('candidatesTokenCount')}",
                file=sys.stderr,
            )
            return None

        parts = cand0.get("content", {}).get("parts", [])
        if not parts:
            print(
                f"   ⚠️ Gemini vrnil prazen content (finishReason={finish_reason}, "
                f"thoughts={usage.get('thoughtsTokenCount')})",
                file=sys.stderr,
            )
            return None
        text = parts[0].get("text", "").strip()
        if not text:
            print(
                f"   ⚠️ Gemini vrnil prazen text (finishReason={finish_reason}, "
                f"thoughts={usage.get('thoughtsTokenCount')}, "
                f"output={usage.get('candidatesTokenCount')})",
                file=sys.stderr,
            )
            return None

        result = _parse_llm_response(text, video_duration)
        if not result:
            return None

        print(f"   🤖 Gemini ({model}) izbral: {result['start']:.1f}-{result['end']:.1f}s", file=sys.stderr)
        print(f"      Razlog: {result.get('reason', '')[:80]}", file=sys.stderr)
        print(f"      Struktura: {result.get('structure', '')[:80]}", file=sys.stderr)
        if result.get("corrected_segments"):
            print(f"      Popravljeni segmenti: {len(result['corrected_segments'])}", file=sys.stderr)

        result["source"] = f"gemini:{model}"
        return result
    except urllib.error.HTTPError as e:
        body = e.read().decode("utf-8", errors="replace")[:500]
        print(f"   ❌ Gemini API HTTP {e.code}: {body}", file=sys.stderr)
        return None
    except Exception as e:
        print(f"   ❌ Gemini analysis failed: {e}", file=sys.stderr)
        return None


def analyze_with_llm(transcript, video_duration, target_duration=30, provider="claude", llm_model=None, filename_hint=None, include_prebuild=False):
    """Glavna funkcija — uporabi izbrano LLM (claude/gemini/auto)."""
    if provider == "gemini":
        model = llm_model or "gemini-3.1-pro-preview"
        return analyze_with_gemini(transcript, video_duration, target_duration, model, filename_hint=filename_hint, include_prebuild=include_prebuild)
    elif provider == "claude":
        model = llm_model or "claude-sonnet-4-6"
        return analyze_with_claude(transcript, video_duration, target_duration, model, filename_hint=filename_hint, include_prebuild=include_prebuild)
    elif provider == "auto":
        # Najprej probaj Claude, fallback na Gemini
        result = analyze_with_claude(transcript, video_duration, target_duration,
                                      llm_model or "claude-sonnet-4-6", filename_hint=filename_hint, include_prebuild=include_prebuild)
        if result:
            return result
        print("   🔄 Claude ni uspel, probam Gemini...", file=sys.stderr)
        return analyze_with_gemini(transcript, video_duration, target_duration,
                                    llm_model or "gemini-3.1-pro-preview", filename_hint=filename_hint, include_prebuild=include_prebuild)
    else:
        print(f"   ⚠️ Neznan LLM provider: {provider}", file=sys.stderr)
        return None



def is_instrumental(transcript, video_duration, threshold=0.1):
    """Detekcija ali je pesem instrumentalna.

    Če je vsota trajanja vokalnih segmentov < threshold * video_duration,
    je pesem instrumentalna.
    """
    if not transcript.get("segments"):
        return True
    vocal_duration = sum(
        s["end"] - s["start"] for s in transcript["segments"]
    )
    ratio = vocal_duration / max(video_duration, 1)
    return bool(ratio < threshold)


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("video", help="Vhod video file")
    ap.add_argument("--lang", default=None, help="ISO 639-1 ali 'auto' (default: auto)")
    ap.add_argument("--model", default="large-v3", help="Whisper model")
    ap.add_argument("--target-duration", type=float, default=30.0)
    ap.add_argument("--max-duration", type=float, default=45.0)
    ap.add_argument("--min-duration", type=float, default=20.0)
    ap.add_argument("--include-prebuild", action="store_true",
                    help="Vključi pre-chorus build-up (privzeto: ne)")
    ap.add_argument("--no-claude", action="store_true",
                    help="Preskoči LLM analizo (uporabi samo lokalno heuristiko)")
    ap.add_argument("--llm-provider", default="claude",
                    choices=["claude", "gemini", "auto"],
                    help="Kateri LLM uporabiti za analizo (default: claude)")
    ap.add_argument("--llm-model", default=None,
                    help="Specifičen model (npr. claude-sonnet-4-6, gemini-3.1-pro-preview)")
    ap.add_argument("--filename-hint", default=None,
                    help="Originalno ime datoteke (Claude lahko prepozna pesem)")
    ap.add_argument("--whisper-provider", default="auto",
                    choices=["auto", "soniox", "elevenlabs", "local", "hybrid", "gemini"],
                    help="STT provider: "
                         "soniox=Soniox stt-async-v4 ($0.10/h, 5-15s, najboljši za NZ, PRIPOROČENO), "
                         "elevenlabs=Scribe ($0.40/h, halucinacije pri NZ), "
                         "gemini=Gemini 3 Pro ($3-5/h, počasen), "
                         "auto=Soniox primary + fallback chain (PRIVZETO)")
    ap.add_argument("--json", action="store_true", help="Output JSON")
    ap.add_argument("--output", help="Path za JSON output")
    args = ap.parse_args()

    video = Path(args.video)
    if not video.exists():
        print(f"❌ Video ne obstaja: {video}", file=sys.stderr)
        sys.exit(1)

    duration = get_video_duration(video)
    print(f"📹 Video: {video.name}, {duration:.1f}s", file=sys.stderr)

    # 1. Extract avdio
    audio = extract_audio(video)

    try:
        # 2. Whisper transcript
        lang = None if args.lang in (None, "auto", "") else args.lang
        # Filename hint pomaga Scribu detektirati jezik (Avseniki → SL, Ben Zucker → DE)
        fname_hint = args.filename_hint or video.stem
        transcript = transcribe_full(
            audio, lang=lang, model_size=args.model,
            provider=args.whisper_provider,
            filename_hint=fname_hint,
        )
        print(f"   Transkripcija: {len(transcript['segments'])} segmentov", file=sys.stderr)

        # 3. Energy profile
        print(f"⚡ Energy profile...", file=sys.stderr)
        energies = compute_energy_profile(audio)
        print(f"   Energy samples: {len(energies)}", file=sys.stderr)

        # 4. Instrumental detection
        instrumental = is_instrumental(transcript, duration)
        print(f"🎵 Instrumentalna: {instrumental}", file=sys.stderr)

        # 5a. PRIMARNO: LLM analiza (razume cel tekst pesmi + popravki)
        claude_result = None
        if not instrumental and not args.no_claude:
            provider = args.llm_provider
            print(f"🤖 Pošiljam transkript {provider}-u za analizo...", file=sys.stderr)
            # Filename hint = original filename brez extension (Claude lahko prepozna pesem)
            fname_hint = args.filename_hint or video.stem
            claude_result = analyze_with_llm(
                transcript, duration, target_duration=args.target_duration,
                provider=provider, llm_model=args.llm_model,
                filename_hint=fname_hint,
                include_prebuild=args.include_prebuild,
            )

        # 5b. Find chorus lokalno (kot fallback ali za score-jev preview)
        if not instrumental:
            chorus = find_chorus(transcript, energies, duration)
        else:
            # Za instrumentalne: najdi sekcijo z najvišjo energijo
            window = args.target_duration
            best_start = 0
            best_avg = -100
            t = 0
            while t + window <= duration:
                avg = avg_energy_in_range(energies, t, t + window)
                if avg > best_avg:
                    best_avg = avg
                    best_start = t
                t += 5  # step 5s
            chorus = {
                "best": {
                    "start": best_start,
                    "end": best_start + window,
                    "duration": window,
                    "text_preview": "(instrumental — energy peak)",
                    "score": 0,
                    "avg_rms": round(best_avg, 2),
                },
                "all_candidates": [],
                "avg_rms_total": round(
                    sum(r for (_, r) in energies) / len(energies) if energies else -30, 2
                ),
            }

        # 6. Clip range — LLM (Claude/Gemini) ima prednost, sicer smart_clip_range fallback.
        # POMEMBNO: spremenljivka se zgodovinsko imenuje claude_result, dejansko pa vsebuje
        # rezultat KATEREGA KOLI LLM-ja (Claude ali Gemini) — glej analyze_with_llm().
        # llm_source npr. "claude:claude-sonnet-4-6" ali "gemini:gemini-3.1-pro-preview".
        if claude_result:
            llm_source = claude_result.get("source", "llm")
            
            # ── HALUCINACIJA HANDLING ──
            # Če je Claude detect-iral halucinacijo (npr. Scribe je vrnil
            # "finančni moduli" namesto pesmi), NE zaupamo izbiri clipa,
            # ker LLM ni mogel locirati pravega refrena.
            if claude_result.get("hallucination_detected"):
                print(f"⚠️ HALUCINACIJA DETECT-ANA — fallback na local heuristic "
                      f"(Scribe transkript ne ustreza zvočnemu vsebini)", file=sys.stderr)
                # Reset claude_result — gremo na local fallback
                clip_range = smart_clip_range(
                    chorus, transcript, duration,
                    target_duration=args.target_duration,
                    max_duration=args.max_duration,
                    min_duration=args.min_duration,
                    include_prebuild=args.include_prebuild,
                )
                clip_range["source"] = "local_fallback_after_hallucination"
                clip_range["reason"] = (
                    "STT halucinacija — local heuristic fallback. "
                    "Refren je iz energy-based detekcije, ne iz transkripta. "
                    + clip_range.get("reason", "")
                )
                claude_result = None  # disable extensions
            else:
                clip_range = {
                    "start": claude_result["start"],
                    "end": claude_result["end"],
                    "duration": claude_result["duration"],
                    "reason": f"{llm_source}: " + claude_result.get("reason", ""),
                    "chorus_text": claude_result.get("chorus_text", ""),
                    "structure": claude_result.get("structure", ""),
                    "source": llm_source,
                }
                # Apply max_duration cap če LLM pretirava
                if clip_range["duration"] > args.max_duration:
                    clip_range["end"] = clip_range["start"] + args.max_duration
                    clip_range["duration"] = args.max_duration
                    clip_range["reason"] += " (capped at max_duration)"
                
                # Apply min_duration floor — če je clip prekratek, podaljšaj
                if clip_range["duration"] < args.min_duration:
                    needed = args.min_duration - clip_range["duration"]
                    new_end = min(clip_range["end"] + needed, duration)
                    actual_extension = new_end - clip_range["end"]
                    clip_range["end"] = new_end
                    clip_range["duration"] = clip_range["end"] - clip_range["start"]
                    clip_range["reason"] += f" (extended +{actual_extension:.1f}s to meet min_duration)"

            # ── EXTEND clip end do naslednje naravne pavze ──
            # LLM pogosto reže točno na zadnji besedi refrena, ampak zadnja
            # beseda ima še "ej-ej-ej" outro / pevec drži zadnji ton 1-3s.
            # Razširimo clip do naslednje >= 1s pavze ali instrumentalnega bridg-a,
            # ampak ne čez max_duration + 5s.
            corrected_segs = claude_result.get("corrected_segments") or transcript["segments"]
            current_end = clip_range["end"]
            extension_limit = min(
                clip_range["start"] + args.max_duration + 5,  # max 5s nad max_duration
                duration  # ne čez celoten audio
            )
            
            # ── EXTEND clip START nazaj če Claude začne sredi besede/segmenta ──
            # Refren se pogosto začne na isti besedi kot v transkriptu, ampak Scribe
            # lahko zazna mejo med segmenti **PO** prvi besedi refrena (npr.
            # "Žena me tepe" — beseda "Žena" v prejšnjem segmentu pri 78.0s,
            # nov segment začne pri 78.3s s "tepe"). To pomeni Claude reže
            # PRED besedo "Žena" → odrezana.
            # 
            # Strategija: če clip start pade SREDI segmenta (ne tik na začetku),
            # razširi nazaj na začetek tega segmenta + 0.2s buffer.
            # ── EXTEND clip START nazaj če Claude začne sredi besede ali tik za njo ──
            # Pesem se pogosto začne na isti besedi v transkriptu, ampak Scribe lahko
            # zazna mejo med segmenti **PO** prvi besedi (npr. "Žena me tepe" — "Žena"
            # je v prejšnjem segmentu pri 76.88-77.70s, novi segment začne 78.30).
            # Claude reže tipično na začetku novega segmenta = odrezana prva beseda.
            #
            # Strategija: **na ravni besed** — najdi besedo katere konec je 
            # blizu clip start (±0.5s) IN preveri ali se lahko ta beseda 
            # "naslanja" na clip (z malo pavze do naslednje besede).
            current_start = clip_range["start"]
            
            # Zberi VSE besede z njihovimi timestampi
            # POMEMBNO: Claude corrected_segments NE vsebuje word-level timestamps,
            # samo segment start/end. Word-level je samo v originalnem Scribe transkriptu.
            # Zato vedno uporabi `transcript["segments"]` ne `corrected_segs`.
            all_words = []
            for seg in transcript.get("segments", []):
                for w in seg.get("words", []):
                    if w.get("start") is not None and w.get("end") is not None:
                        all_words.append({
                            "start": float(w["start"]),
                            "end": float(w["end"]),
                            "text": w.get("text", ""),
                        })
            
            if all_words:
                # Najdi "rob" — beseda kjer končanje zelo blizu clip start
                # ALI clip start je sredi besede (besedo bi odrezali)
                # ALI prejšnje besede so del istega govora pred clip start
                
                # Strategija: poišči besedo PRED clip start, nato razširi nazaj
                # za **1-2 besedi** (ne celo frazo - to bi zajelo prejšnji verz).
                # Kombiniraj z amplitude defense (Layer 3) ki dodatno doda buffer.
                MAX_LOOKBACK_WORDS = 2  # max 2 besedi nazaj
                
                for i, w in enumerate(all_words):
                    # Beseda zaobsega clip start (clip reže sredi besede)
                    if w["start"] < current_start < w["end"]:
                        # Razširi nazaj na začetek te besede in največ MAX_LOOKBACK_WORDS predhodnih
                        anchor_idx = i
                        for j in range(i, max(0, i - MAX_LOOKBACK_WORDS), -1):
                            prev = all_words[j - 1]
                            curr = all_words[j]
                            gap = curr["start"] - prev["end"]
                            if gap >= 0.5:
                                break
                            anchor_idx = j - 1
                        new_start = max(0, all_words[anchor_idx]["start"])  # NI buffer-ja
                        captured = " ".join(w2["text"].strip() for w2 in all_words[anchor_idx:i+1])
                        print(f"   🎵 Razširim clip začetek {current_start:.2f}s → {new_start:.2f}s "
                              f"(clip sredi besede; ujamem '{captured}')", file=sys.stderr)
                        current_start = new_start
                        break
                    # Beseda končana TIK pred clip start (do 0.5s pred)
                    if 0 < (current_start - w["end"]) <= 0.5:
                        # Preveri naslednjo besedo
                        next_w = all_words[i + 1] if i + 1 < len(all_words) else None
                        if next_w and next_w["start"] >= current_start - 0.1:
                            # Najdi anchor: do MAX_LOOKBACK_WORDS nazaj
                            anchor_idx = i
                            for j in range(i, max(0, i - MAX_LOOKBACK_WORDS), -1):
                                prev = all_words[j - 1]
                                curr = all_words[j]
                                gap = curr["start"] - prev["end"]
                                if gap >= 0.5:
                                    break
                                anchor_idx = j - 1
                            new_start = max(0, all_words[anchor_idx]["start"])  # NI buffer-ja
                            captured = " ".join(w2["text"].strip() for w2 in all_words[anchor_idx:i+1])
                            print(f"   🎵 Razširim clip začetek {current_start:.2f}s → {new_start:.2f}s "
                                  f"(beseda '{w['text'].strip()}' tik pred clip start; "
                                  f"ujamem celo frazo '{captured}')", file=sys.stderr)
                            current_start = new_start
                            break
            else:
                # Fallback: če ni word-level (npr. local Whisper), uporabi segmente kot prej
                for seg in corrected_segs:
                    seg_start = float(seg.get("start", 0))
                    seg_end = float(seg.get("end", 0))
                    if seg_start < current_start < seg_end:
                        new_start = max(0, current_start - 0.5)
                        print(f"   🎵 Razširim clip začetek {current_start:.2f}s → {new_start:.2f}s "
                              f"(brez word-level, fallback -0.5s)", file=sys.stderr)
                        current_start = new_start
                        break
            
            if current_start < clip_range["start"]:
                clip_range["start"] = round(current_start, 2)
                clip_range["duration"] = round(clip_range["end"] - current_start, 2)
                clip_range["reason"] += f" (start extended back)"
            
            # ── SLOJ 3: AUDIO AMPLITUDE CHECK na samem začetku clipa ──
            # Tudi po word-level extension lahko clip začne sredi vokala (npr. če
            # Scribe ni zaznal besede). Kot zadnja obramba: preveri RMS audio
            # amplitudo v prvih 100ms clipa. Če je > silence threshold = vokal je
            # že tam, dodaj 0.5s buffer nazaj.
            try:
                import subprocess as _sp
                # ffmpeg lahko prebere kratek segment in vrne RMS volume
                probe_start = clip_range["start"]
                probe_dur = 0.15  # prvih 150ms
                if probe_start >= 0.5:  # samo če imamo prostor za buffer
                    cmd_probe = [
                        "ffmpeg", "-hide_banner", "-loglevel", "error",
                        "-ss", str(probe_start), "-t", str(probe_dur),
                        "-i", str(args.video),
                        "-af", "volumedetect",
                        "-f", "null", "-"
                    ]
                    pr = _sp.run(cmd_probe, capture_output=True, text=True, timeout=10)
                    output = pr.stderr or ""
                    # Iščemo "mean_volume: -XX.X dB"
                    import re as _re_amp
                    m = _re_amp.search(r'mean_volume:\s*(-?\d+\.?\d*)\s*dB', output)
                    if m:
                        mean_db = float(m.group(1))
                        # Silence threshold: pod -40 dB = tihota
                        # Vokal/glasba je običajno -30 do -10 dB
                        if mean_db > -35:
                            # Audio je že "glasen" na začetku clipa = vokal/glasba
                            # Dodaj 0.5s buffer nazaj (varno, ne prepogosto)
                            old_start = clip_range["start"]
                            new_start = max(0, old_start - 0.5)
                            if new_start < old_start:
                                print(f"   🎵 Audio amplitude check: prvih {probe_dur}s "
                                      f"ima mean_volume {mean_db:.1f} dB (> -35 dB = vokal/glasba). "
                                      f"Razširim clip {old_start:.2f}s → {new_start:.2f}s.", file=sys.stderr)
                                clip_range["start"] = round(new_start, 2)
                                clip_range["duration"] = round(clip_range["end"] - new_start, 2)
                                clip_range["reason"] += " (amplitude defense -0.5s)"
                        else:
                            print(f"   🎵 Audio amplitude check: prvih {probe_dur}s "
                                  f"ima mean_volume {mean_db:.1f} dB (≤ -35 dB = tiho). OK.", file=sys.stderr)
            except Exception as _e:
                print(f"   ⚠️ Audio amplitude check skipped: {_e}", file=sys.stderr)
            
            # Najdi vse segmente ki se začnejo PO trenutnem clip end
            # STROŽJA pravila: ne podaljšuj v naslednji refren / verz / instrumental.
            # Razširjamo SAMO če zadnji segment se prekriva s clip (klesti iz njega) ALI
            # če je naslednji segment KRATEK (< 2s) IN vsebuje samo outro fillerje
            # (la la, oh, yeah, ej, ja, ah, na, hey itd.).
            
            # Definiraj outro filler regex (multi-jezikovno)
            import re as _re
            OUTRO_FILLER_RE = _re.compile(
                r'^[\s\-,.!?]*'
                r'((?:la|na|oh|ah|eh|ej|aj|ja|hey|yeah|yo|ho|wo|hu|mm|nn|uu|oo|aa|ee|ii)'
                r'[\s\-,.!?]*)+'
                r'[\s\-,.!?]*$',
                _re.IGNORECASE
            )
            # Hard cap: ne razširjaj več kot 3s nad původne clip end
            original_clip_end = clip_range["end"]
            soft_extension_limit = min(original_clip_end + 3.0, extension_limit)
            
            for seg in corrected_segs:
                seg_start = float(seg.get("start", 0))
                seg_end = float(seg.get("end", 0))
                seg_text = seg.get("text", "").strip()
                
                # Segment se prekriva s clip end (zadnji segment refrena, ki ni zaključen)
                if seg_start <= current_end:
                    if seg_end > current_end and seg_end <= soft_extension_limit:
                        new_end = min(seg_end + 0.3, soft_extension_limit)
                        if new_end > current_end:
                            print(f"   🎵 Podaljšam clip {current_end:.1f}s → {new_end:.1f}s "
                                  f"(zadnji segment refrena se zaključi)", file=sys.stderr)
                            current_end = new_end
                else:
                    # Segment začne PO clip end — preveri ali je outro filler
                    pause = seg_start - current_end
                    
                    # Predaleč → ustavi se
                    if pause >= 0.7:
                        break
                    # Predolg segment = nov verz/refren, ne dodaj
                    if (seg_end - seg_start) > 2.5:
                        break
                    # Preveri vsebino — če ni samo outro fillerji, NE dodaj
                    if not OUTRO_FILLER_RE.match(seg_text):
                        # Ni filler → verjetno nov refren/verz/post-chorus
                        break
                    
                    # OK, je outro filler — dodaj
                    new_end = min(seg_end + 0.2, soft_extension_limit)
                    if new_end > current_end:
                        print(f"   🎵 Podaljšam clip {current_end:.1f}s → {new_end:.1f}s "
                              f"(outro filler '{seg_text[:40]}')", file=sys.stderr)
                        current_end = new_end
                    else:
                        break
            
            if current_end > clip_range["end"]:
                clip_range["end"] = round(current_end, 2)
                clip_range["duration"] = round(current_end - clip_range["start"], 2)
                clip_range["reason"] += f" (extended to natural pause)"
        else:
            clip_range = smart_clip_range(
                chorus, transcript, duration,
                target_duration=args.target_duration,
                max_duration=args.max_duration,
                min_duration=args.min_duration,
                include_prebuild=args.include_prebuild,
            )
            clip_range["source"] = "local_heuristic"
        print(f"✂  Clip range: {clip_range['start']:.1f}s - {clip_range['end']:.1f}s "
              f"(duration: {clip_range['duration']}s, source: {clip_range.get('source')})",
              file=sys.stderr)

        # Če Claude je vrnil popravljene segmente, jih uporabi (boljši za podnapise)
        if claude_result and claude_result.get("corrected_segments"):
            corrected = claude_result["corrected_segments"]
            # Ohrani word-level timing iz originala, posodobi samo text
            orig_by_start = {round(s["start"], 1): s for s in transcript["segments"]}
            new_segments = []
            for cs in corrected:
                try:
                    cs_start = float(cs["start"])
                    cs_end = float(cs["end"])
                    cs_text = str(cs["text"]).strip()
                except (KeyError, ValueError, TypeError):
                    continue
                # Najdi originalni segment z istim start (ali blizu) za word-level timing
                orig = orig_by_start.get(round(cs_start, 1))
                if not orig:
                    # Najdi najbližji
                    closest_diff = 999
                    for s in transcript["segments"]:
                        diff = abs(s["start"] - cs_start)
                        if diff < closest_diff and diff < 1.0:
                            closest_diff = diff
                            orig = s
                new_segments.append({
                    "start": cs_start,
                    "end": cs_end,
                    "text": cs_text,
                    # Word-level timing ne moremo posodabljati ker Claude ne vrača besede,
                    # ampak ohranimo če imamo
                    "words": orig.get("words", []) if orig else [],
                })
            transcript["segments"] = new_segments
            transcript["claude_corrected"] = True  # ohranimo ime ključa zaradi backward compat
            # Posodobi tudi jezik če LLM je drugačnega mnenja
            if claude_result.get("language") and claude_result["language"] != transcript["language"]:
                print(f"   ✏️ LLM je popravil jezik: {transcript['language']} → {claude_result['language']}", file=sys.stderr)
                transcript["language"] = claude_result["language"]
            llm_label = claude_result.get("source", "LLM")
            print(f"   ✏️ Whisper segmenti zamenjani z {llm_label} popravljenimi ({len(new_segments)})", file=sys.stderr)

        # 7. Fade params (lahko razširi clip end če konča sredi vokala)
        fade = detect_audio_fade(clip_range, transcript, video_duration=duration)
        print(f"🎚  Fade: in={fade['fade_in']}s, out={fade['fade_out']}s", file=sys.stderr)

        # Če fade detection razširi end (ker clip konča sredi vokala), apply
        if fade.get("extended_end") and fade["extended_end"] > clip_range["end"]:
            old_end = clip_range["end"]
            new_end = min(fade["extended_end"], clip_range["start"] + args.max_duration)
            clip_range["end"] = round(new_end, 2)
            clip_range["duration"] = round(new_end - clip_range["start"], 2)
            print(f"   ↳ Razširjen za {new_end - old_end:.1f}s (zaključek besedila)",
                  file=sys.stderr)

        result = {
            "video": str(video),
            "video_duration": duration,
            "language": transcript["language"],
            "language_probability": transcript["language_probability"],
            "instrumental": instrumental,
            "transcript": transcript,
            "chorus": chorus,
            "clip_range": clip_range,
            "fade": fade,
            "claude_used": claude_result is not None,
            "claude_corrected_text": bool(claude_result and claude_result.get("corrected_segments")),
        }

        if args.output:
            with open(args.output, "w", encoding="utf-8") as f:
                json.dump(result, f, ensure_ascii=False, indent=2)
            print(f"💾 Saved: {args.output}", file=sys.stderr)

        if args.json:
            print(json.dumps(result, ensure_ascii=False))

    finally:
        try:
            os.unlink(audio)
        except Exception:
            pass


if __name__ == "__main__":
    main()