Trust Gemini timestamps over unreliable ffprobe duration

The container duration probe under-reported a ~5min video as 209s
(bad WhatsApp/edited MP4 metadata), causing a valid 230-329s clip to be
skipped. Stop using the probe to skip/clamp or to constrain Gemini's
prompt; trust Gemini's timestamps (it watched the real file) and let
-c copy fail gracefully on genuinely out-of-range segments. Probe is now
diagnostic-only. Raise clip cap to 300s so long steps aren't truncated.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

main.py

@@ -183,8 +183,9 @@ def _storage_signed_url(path: str, minutes: int = 120) -> str:
         return ""
 
 
-# Maximum length (seconds) of a tutorial clip we will cut. Long answers stay text.
-MAX_CLIP_SECONDS = 120
+# Maximum length (seconds) of a tutorial clip we will cut. Generous so full steps
+# aren't truncated; stream-copied clips stay well under Twilio's ~16MB media limit.
+MAX_CLIP_SECONDS = 300
 # Padding added to the end of a clip so the demonstrated step is not cut off abruptly.
 CLIP_TAIL_PAD    = 2
 
@@ -1242,29 +1243,6 @@ def tutorial_ingest():
     logger.info("[tutorial] STEP 1/5: received '%s' (%.2f MB, mime=%s)",
                 video_title, len(video_bytes) / 1024 / 1024, mime_type)
 
-    # Probe the real duration up front so we can constrain Gemini's timestamps to it —
-    # otherwise the model sometimes invents segments past the end of the video.
-    vid_dur_hint = 0.0
-    try:
-        with tempfile.NamedTemporaryFile(suffix=ext, delete=False) as _pt:
-            _pt.write(video_bytes)
-            _probe_path = _pt.name
-        vid_dur_hint = _video_duration(_probe_path)
-        logger.info("[tutorial] probed duration %.1fs, codec %s",
-                    vid_dur_hint, _video_codec(_probe_path))
-        os.remove(_probe_path)
-    except Exception as e:
-        logger.warning("[tutorial] pre-probe failed: %s", e)
-
-    extraction_prompt = TUTORIAL_VIDEO_PROMPT
-    if vid_dur_hint and vid_dur_hint > 0:
-        extraction_prompt += (
-            f"\n\nIMPORTANT: This video is exactly {int(vid_dur_hint)} seconds long. "
-            f"Every timestamp_start and timestamp_end MUST be between 0 and {int(vid_dur_hint)}. "
-            f"Do not output any timestamp greater than {int(vid_dur_hint)}, and only describe "
-            f"steps that actually occur within the video's length."
-        )
-
     # Upload to Gemini Files API and wait for processing
     logger.info("[tutorial] STEP 2/5: uploading to Gemini Files API + waiting for ACTIVE...")
     gemini_file = _upload_video_to_gemini(video_bytes, mime_type, display_name=video_title)
@@ -1280,7 +1258,7 @@ def tutorial_ingest():
         t_extract = time.time()
         resp = _gemini_client.models.generate_content(
             model=GEMINI_MODEL,
-            contents=[gemini_file, extraction_prompt],
+            contents=[gemini_file, TUTORIAL_VIDEO_PROMPT],
             config=genai_types.GenerateContentConfig(
                 response_mime_type="application/json",
                 http_options=genai_types.HttpOptions(timeout=GEMINI_VIDEO_TIMEOUT_MS),
@@ -1338,9 +1316,12 @@ def tutorial_ingest():
             with tempfile.NamedTemporaryFile(suffix=ext, delete=False) as src_tmp:
                 src_tmp.write(video_bytes)
                 src_path = src_tmp.name
-            vid_dur = _video_duration(src_path)
-            logger.info("[tutorial] wrote source video to temp: %s (duration %.1fs, codec %s)",
-                        src_path, vid_dur, _video_codec(src_path))
+            # Probe is logged for diagnostics only — container metadata is sometimes
+            # wrong (esp. WhatsApp/edited MP4s), so we do NOT use it to skip segments.
+            # We trust Gemini's timestamps (it watched the real file) and let ffmpeg
+            # copy fail gracefully if a segment is genuinely past EOF.
+            logger.info("[tutorial] wrote source video to temp: %s (reported duration %.1fs, codec %s)",
+                        src_path, _video_duration(src_path), _video_codec(src_path))
             for idx, a in enumerate(articles, 1):
                 start = a.get("timestamp_start", 0)
                 end   = a.get("timestamp_end", 0)
@@ -1351,7 +1332,7 @@ def tutorial_ingest():
                 out_path  = os.path.join(tempfile.gettempdir(), f"clip_{uuid.uuid4().hex}.mp4")
                 clip_path = f"kb_clips/{uuid.uuid4().hex}.mp4"
                 logger.info("[tutorial] clip %d/%d: cropping %ss–%ss...", idx, len(articles), start, end)
-                if _crop_video_segment(src_path, start, end, out_path, video_duration=vid_dur):
+                if _crop_video_segment(src_path, start, end, out_path):
                     try:
                         clip_bytes = os.path.getsize(out_path)
                         logger.info("[tutorial] clip %d/%d: uploading %d bytes to %s...",