refactor: extract provider logic into modular registry

Each AI provider now has its own module implementing a ProviderModule
interface with createGateway, getAuthHeaders, and fetchModels methods.
This replaces the monolithic switch statements in ai-provider.ts and
models.post.ts with a clean registry pattern.

server/utils/providers/vllm.ts

@@ -0,0 +1,42 @@
+import { Ok } from '~~/types/result';
+import type { ProviderModule } from '.';
+import { createOpenAICompatible } from '@ai-sdk/openai-compatible';
+
+export default {
+    id: 'vllm',
+    baseUrl: '',
+    modelsEndpoint: '/models',
+
+    createGateway({ apiKey, baseURL }) {
+        return Ok({
+            gateway: createOpenAICompatible({ name: 'vLLM', apiKey, baseURL, includeUsage: true }),
+            streamTransformer: undefined,
+            textTransformer: undefined,
+        });
+    },
+
+    getAuthHeaders(apiKey) {
+        return apiKey ? { Authorization: `Bearer ${apiKey}` } : {};
+    },
+
+    async fetchModels({ baseURL, apiKey }) {
+        const res = await fetch(`${baseURL}/models`, {
+            method: 'GET',
+            headers: { Authorization: `Bearer ${apiKey}` },
+        });
+        const data = await res.json();
+
+        if (!res.ok) {
+            throw createError({ statusCode: res.status, message: JSON.stringify(data) });
+        }
+
+        return data.data.map((model: any) => ({
+            id: model.id,
+            attributes: {
+                inputModalities: ['text'],
+                outputModalities: ['text'],
+                contextWindow: model.max_model_len,
+            }
+        }));
+    },
+} satisfies ProviderModule;