chore: log where each answer's time goes — rewrite, search, first token, total

Numbers only, one line per streamed question, so "is the chat slow?" can be
answered from the log instead of from a feeling. Nothing about the request
is logged beyond the counts.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU

src/routes/clinicalAssistant.js

@@ -423,6 +423,7 @@ router.post('/clinical-assistant/chat/stream', async function(req, res) {
 
     var toolset = assistantToolset(prepared);
     if (prepared.delegateVision) sendEvent('status', { message: 'Looking at your image…' });
+    var firstTokenAt = 0;
     var ai = await callAIStream(prepared.messages, assistantGenerationOptions({
       model: prepared.chatModel || undefined,
       temperature: 0.15,
@@ -430,8 +431,13 @@ router.post('/clinical-assistant/chat/stream', async function(req, res) {
       maxTokens: 2600,
       images: toolset.images
     }), function(delta) {
+      if (!firstTokenAt) firstTokenAt = Date.now();
       sendEvent('token', { token: delta });
     });
+    var t = prepared.timing || {};
+    console.info('[clinical-assistant] timing ms: rewrite=' + (t.rewriteMs || 0) + ' search=' + (t.searchMs || 0) +
+      ' first_token=' + (firstTokenAt ? firstTokenAt - started : 0) + ' total=' + (Date.now() - started) +
+      ' sources=' + (prepared.sources || []).length + ' model=' + (ai.model || prepared.chatModel || ''));
     ai = await visionTool.dispatch(ai, {
       images: prepared.images, visionModel: prepared.visionModel, callAI: callAI,
       messages: prepared.messages,
@@ -573,15 +579,22 @@ async function prepareAssistantChat(body) {
   var includeContext = body.includeContext !== false;
   var showSources = await showSourcesEnabled();
 
+  // Phase timings, numbers only, so "is it slow?" can be answered from the
+  // log rather than from a feeling: how long the rewrite took, how long the
+  // library search took, and (in the stream route) when the first token came.
+  var timing = { rewriteMs: 0, searchMs: 0 };
+  var phase = Date.now();
   var searchQuery = await rewriteSearchQuery(message, history, chatModel).catch(function(e) {
     console.warn('[clinical-assistant] query rewrite skipped:', e.message);
     return message;
   });
+  timing.rewriteMs = Date.now() - phase; phase = Date.now();
   var searchResponse = await semanticSearch(searchQuery, {
     limit: searchLimit,
     includeContext: includeContext,
     contextChars: contextChars
   });
+  timing.searchMs = Date.now() - phase;
   // Retrieval is text-only. The multimodal path called nc_multimodal_search
   // against a second hardcoded collection with an embedding service that was
   // never deployed, so it only ever logged "multimodal search skipped".
@@ -592,7 +605,7 @@ async function prepareAssistantChat(body) {
   var sources = dedupeSources(rawResults).slice(0, searchLimit);
 
   if (sources.length === 0) {
-    return { direct: {
+    return { timing: timing, direct: {
       success: true,
       answer: 'I could not find a clear match for that question. Please try a more specific clinical term, diagnosis, medication, age group, or textbook topic.',
       sources: [],
@@ -603,6 +616,7 @@ async function prepareAssistantChat(body) {
 
   var context = formatSourcesForPrompt(sources);
   return {
+    timing: timing,
     message: message,
     showSources: showSources,
     images: images,