feat(podcasts): add YouTube podcast source with transcript-first discovery

New "podcasts" source that discovers podcast content by scanning transcripts
from LLM-resolved YouTube channels. Finds content invisible to title-based
search — Acquired's "The NFL" episode mentions Taylor Swift 18x, ESPN 117x,
Netflix 102x, none in the title.

Architecture:
- LLM resolves 6-12 podcast channel @handles per topic
- Engine fetches recent episodes via yt-dlp (no video download)
- Downloads auto-captions and greps for topic keywords
- Episodes with 5+ mentions become podcast results with highlights
- Runs in parallel, ~15-20s latency, invisible in 3-min research run

Pipeline integration:
- New source module: scripts/lib/podcast_yt.py
- Registered in pipeline, normalizer, signals, planner, render
- CLI flag: --podcast-channels=AcquiredFM,lexfridman,...
- SOURCE_QUALITY: 0.88 (above YouTube's 0.85)
- Opt-in via INCLUDE_SOURCES=podcasts or --search=podcasts

Zero new API keys. Zero new dependencies. Reuses yt-dlp + transcript pipeline.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

scripts/lib/planner.py

@@ -71,6 +71,7 @@ SOURCE_CAPABILITIES = {
     "github": {"discussion", "link"},
     "grounding": {"web", "reference", "link"},
     "perplexity": {"web", "reference", "analysis"},
+    "podcasts": {"discussion", "video_longform", "expert"},
 }
 DEFAULT_INTENT_CAPABILITIES = {
     "comparison": {"discussion", "video", "web", "reference", "social", "link", "market"},