add vision analysis service with prompt versioning, response schema, and structured JSON output

src/services/modelService.js

@@ -104,7 +104,7 @@ export function createModelService(config, overrides = {}) {
         }
     };
 
-    const readStreamingReply = async (response, onChunk = null, onThinking = null, touch = () => { }, signal = null) => {
+    const readStreamingReply = async (response, onChunk = null, onThinking = null, touch = () => { }, signal = null, onRawReply = null) => {
         if (!response.body) {
             throw new ModelServiceError("MODEL_BAD_RESPONSE", "Model backend returned no body.", 502);
         }
@@ -161,6 +161,7 @@ export function createModelService(config, overrides = {}) {
                 consume(buffer.trim());
             }
             if (!completed) throw new ModelServiceError("MODEL_INTERRUPTED", "Model response was interrupted.", 502);
+            onRawReply?.(fullReply);
             return fullReply.trim();
         } finally {
             signal?.removeEventListener("abort", cancel);
@@ -187,6 +188,18 @@ export function createModelService(config, overrides = {}) {
     };
 
     return {
+        getRequestMetadata(model, generation = {}, reasoning = false) {
+            const body = buildRequestBody(model, [], generation, reasoning);
+            return {
+                model: body.model,
+                think: body.think,
+                stream: body.stream,
+                temperature: body.options?.temperature ?? null,
+                numCtx: body.options?.num_ctx ?? null,
+                numPredict: body.options?.num_predict ?? null,
+                keepAlive: body.keep_alive
+            };
+        },
         async generateReply(model, messagesPayload, options = {}) {
             const state = ensureModelState(model);
             state.activeRequests += 1;
@@ -206,7 +219,8 @@ export function createModelService(config, overrides = {}) {
                     }
 
                     touch();
-                    return readStreamingReply(response, null, options.reasoning === true && supportsReasoning(model) ? options.onThinking : null, touch, signal);
+                    return readStreamingReply(response, null, options.reasoning === true && supportsReasoning(model) ? options.onThinking : null,
+                        touch, signal, options.onRawReply);
                 }, options.signal);
                 lastHealth = { ok: true, checkedAt: new Date().toISOString(), error: null, latencyMs: Date.now() - startedAt };
                 return fullReply;