fix: use fast model for story selection, gemma4 for summaries only

gemma4:26b under format:json ignores the bare-array instruction and returns a
reasoning object ({"thought": ...}), which broke Phase-1 selection (Unterminated
string -> silent fallback to first-5). Selection now uses ollama_select_model
(default qwen2.5:1.5b); summaries stay on gemma4:26b. Also tolerate dict-shaped
selection responses defensively.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

config.example.json

@@ -4,6 +4,7 @@
     "freshrss_api_key": "your-api-password",
     "ollama_url": "http://your-ollama-ip:11434",
     "ollama_model": "gemma4:26b",
+    "ollama_select_model": "qwen2.5:1.5b",
     "yourls_url": "http://your-yourls-ip/yourls-api.php",
     "yourls_user": "yourls-username",
     "yourls_pass": "yourls-password",