feat(confidence-engine): constrain reconstruction chat output

docs/current-handoff.md

@@ -82,7 +82,9 @@
 
 - Zod upgraded from 3.25.76 to 4.5.4. Reconstruction and direct Zod-owner suites pass, and native `z.toJSONSchema()` converts the production reconstruction schema with required closed enum constraints.
 - Zod-3-specific error-message assertions are version-independent while retaining semantic validation evidence. Production schema semantics are unchanged and no third-party JSON-schema conversion dependency was added.
-- Next unknown: whether Ollama/Qwen accepts and obeys the generated schema through `/api/chat` format.
+- Reconstruction `/api/chat` now sends `z.toJSONSchema(reconstructionV2Schema)` rather than generic `format: "json"`; the same canonical Zod contract constrains generation and remains the post-hoc validator.
+- The deterministic provider-boundary test passes 2/2 under Node. A direct configured Ollama/Qwen `/api/chat` call accepted and obeyed the full schema in one call; `/api/generate`, chat detection, fallback, prompt, temperature, retries, and schema semantics are unchanged.
+- Production repeatability remains untested after this change. Next restart point: a small repeated `/api/cases/start` stability observation using the fixed manufacturing scenario.
 
 ## Current product architecture
 

lib/llm/provider.js

@@ -1,3 +1,6 @@
+import { z } from "zod";
+import { reconstructionV2Schema } from "../reconstruction/schema.js";
+
 /**
  * Provider abstraction — the app calls getProvider() which returns an object
  * with a generateReconstruction(scenario, modelName) method.
@@ -59,6 +62,7 @@ function recoverJson(raw) {
 }
 
 let _chatSupported = null;
+const reconstructionJsonSchema = z.toJSONSchema(reconstructionV2Schema);
 
 async function detectChatSupport(baseUrl, modelName) {
   if (_chatSupported !== null) return _chatSupported;
@@ -113,7 +117,7 @@ class OllamaLlmProvider {
     } catch { /* failed silently — defaults to false */ }
 
     // ================================================================
-    // Step 2: Try /api/chat if supported and format:json works
+    // Step 2: Try /api/chat if supported with the reconstruction schema
     // ================================================================
     if (chatSupported) {
       try {
@@ -128,7 +132,7 @@ class OllamaLlmProvider {
             model: modelName,
             messages: [{ role: "user", content: prompt }],
             stream: false,
-            format: "json",
+            format: reconstructionJsonSchema,
           }),
           signal: controller.signal,
         });

tests/llm/provider.test.js

@@ -23,6 +23,17 @@ describe("OllamaLlmProvider chat capability detection", () => {
       expect(fetchSpy.mock.calls[0][0]).toBe("http://ollama.test/api/chat");
       expect(fetchSpy.mock.calls[1][0]).toBe("http://ollama.test/api/chat");
       expect(fetchSpy.mock.calls[1][0]).not.toContain("/api/generate");
+      const chatRequest = JSON.parse(fetchSpy.mock.calls[1][1].body);
+      expect(chatRequest).toMatchObject({
+        model: "configured-model",
+        messages: [{ role: "user", content: "prompt" }],
+        stream: false,
+      });
+      expect(chatRequest.format).toBeTypeOf("object");
+      expect(chatRequest.format).not.toBe("json");
+      const formatText = JSON.stringify(chatRequest.format);
+      expect(formatText).toContain("relationship");
+      expect(formatText).toContain("evidenceType");
       expect(result).toMatchObject({
         response: {},
         providerApiPath: "/api/chat",