Devon own stem/track/channel, per-category LLM routing, settings UI cleanup

Audio:
- Devon gets own stem, Reaper track (Input 17), and configurable channel
- play_caller_audio accepts stem_name + channel_override params
- Reaper script checks 4 voice tracks (Host, Devon, Live Caller, AI Caller)
- postprod.py includes devon stem in gap detection

Cost optimization:
- Per-category model routing: Sonnet for caller dialog, Gemini Flash for everything else
- Estimated 65% cost reduction ($4.32 → ~$1.50/show)
- Category models configurable from settings UI

Frontend:
- Settings panel: clean routing grid for output channels, model routing grid for LLM categories
- Devon channel added to audio routing
- Share icon SVG fill fix (currentColor)
- Website homepage iterations

Publishing:
- Revert Castopod API workaround (API re-enabled)
- Fix container media path

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

postprod.py

@@ -17,7 +17,7 @@ from pathlib import Path
 import numpy as np
 import soundfile as sf
 
-STEM_NAMES = ["host", "caller", "music", "sfx", "ads", "idents"]
+STEM_NAMES = ["host", "caller", "devon", "music", "sfx", "ads", "idents"]
 
 
 def load_stems(stems_dir: Path) -> tuple[dict[str, np.ndarray], int]:
@@ -69,7 +69,7 @@ def remove_gaps(stems: dict[str, np.ndarray], sr: int,
 
     # Detect gaps in everything except music (which always plays).
     # This catches TTS latency gaps while protecting ad breaks and SFX transitions.
-    content = stems["host"] + stems["caller"] + stems["sfx"] + stems["ads"] + stems["idents"]
+    content = stems["host"] + stems["caller"] + stems["devon"] + stems["sfx"] + stems["ads"] + stems["idents"]
     rms = compute_rms(content, window_samples)
 
     # Threshold: percentile-based to sit above the mic noise floor