fix(parser): handle thinking models in recipe detection

Increase max_tokens from 10 to 1024 for detection so thinking
models have room to reason. Also fall back to reasoning_content
if content is empty, since some local models (e.g. Gemma 4
thinking variants) put their answer there.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

src/lib/server/parser.ts

@@ -49,11 +49,17 @@ export async function detectRecipe(text: string): Promise<boolean> {
 					content: `Does this text contain a recipe?\n\n${text}`
 				}
 			],
-			max_tokens: 10,
+			// 1024 gives thinking models room to reason before answering
+			max_tokens: 1024,
 			temperature: 0
 		});
 
-		const detectionResult = detectionResponse.choices[0].message.content?.toLowerCase() ?? '';
+		const msg = detectionResponse.choices[0].message;
+		// Some local models (e.g. Gemma thinking variants) return the answer in
+		// reasoning_content instead of content when max_tokens is tight.
+		// eslint-disable-next-line @typescript-eslint/no-explicit-any
+		const reasoning: string = (msg as any).reasoning_content ?? '';
+		const detectionResult = (msg.content ?? reasoning).toLowerCase();
 		console.log('[LLM] Detection response:', detectionResult);
 
 		return detectionResult.includes('yes');