fix(youtube): pass full transcripts to LLM instead of truncating to 200 chars

TRANSCRIPT_MAX_WORDS raised from 500 to 5000 so the LLM gets the full
content of most videos (up to ~25 minutes). Removed the second 200-char
truncation in render.py that was reducing transcripts to a single sentence
before the judge agent ever saw them.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

scripts/lib/render.py

@@ -258,10 +258,7 @@ def render_compact(report: schema.Report, limit: int = 15, missing_keys: str = "
             lines.append(f"  {item.title}")
             lines.append(f"  {item.url}")
             if item.transcript_snippet:
-                snippet = item.transcript_snippet[:200]
+                lines.append(f"  Transcript: {item.transcript_snippet}")
-                if len(item.transcript_snippet) > 200:
-                    snippet += "..."
-                lines.append(f"  Transcript: {snippet}")
             lines.append(f"  *{item.why_relevant}*")
             lines.append("")
 

scripts/lib/youtube_yt.py

@@ -33,7 +33,7 @@ TRANSCRIPT_LIMITS = {
 }
 
 # Max words to keep from each transcript
-TRANSCRIPT_MAX_WORDS = 500
+TRANSCRIPT_MAX_WORDS = 5000
 
 from .relevance import token_overlap_relevance as _compute_relevance