fix(decision-ollama): use Ollama's native JSON mode without thinking

The 9B model spent its whole 4000-token budget on reasoning and came back
model_incomplete. Call /api/chat with format=json, think=false and num_predict,
and read prompt_eval_count/eval_count/done_reason. Structured extraction no
longer needs chain-of-thought.

src/document-workbench/decision-ollama-client.ts

@@ -100,7 +100,7 @@ export class DecisionOllamaGatewayClient {
 
     let response: Response;
     try {
-      response = await this.fetchImpl(`${this.baseUrl}/v1/chat/completions`, {
+      response = await this.fetchImpl(`${this.baseUrl}/api/chat`, {
         method: 'POST',
         redirect: 'error',
         signal: AbortSignal.timeout(this.options.timeoutMs ?? 120_000),
@@ -113,9 +113,20 @@ export class DecisionOllamaGatewayClient {
         body: JSON.stringify({
           model: this.options.model,
           messages,
-          max_tokens: admission.maxOutputTokens,
-          temperature: 0,
           stream: false,
+          // Structured extraction: JSON mode, no chain-of-thought output, and a
+          // hard output budget. A thinking model otherwise spends the whole
+          // budget on reasoning and returns model_incomplete.
+          format: 'json',
+          think: false,
+          options: {
+            temperature: 0,
+            num_predict: admission.maxOutputTokens,
+            num_ctx: Math.max(
+              4096,
+              Math.min(131072, admission.maxInputTokens + admission.maxOutputTokens + 1024),
+            ),
+          },
         }),
       });
     } catch {
@@ -131,13 +142,15 @@ export class DecisionOllamaGatewayClient {
     }
     const result = decoded as {
       model?: unknown;
-      choices?: Array<{ message?: { content?: unknown }; finish_reason?: unknown }>;
-      usage?: { prompt_tokens?: unknown; completion_tokens?: unknown; total_tokens?: unknown };
+      message?: { content?: unknown };
+      done?: unknown;
+      done_reason?: unknown;
+      prompt_eval_count?: unknown;
+      eval_count?: unknown;
     } | null;
-    const content = result?.choices?.[0]?.message?.content;
-    const promptTokens = result?.usage?.prompt_tokens;
-    const completionTokens = result?.usage?.completion_tokens;
-    const totalTokens = result?.usage?.total_tokens;
+    const content = result?.message?.content;
+    const promptTokens = result?.prompt_eval_count;
+    const completionTokens = result?.eval_count;
     if (
       !result ||
       typeof content !== 'string' ||
@@ -145,16 +158,15 @@ export class DecisionOllamaGatewayClient {
       !result.model.trim() ||
       !Number.isSafeInteger(promptTokens) ||
       !Number.isSafeInteger(completionTokens) ||
-      !Number.isSafeInteger(totalTokens) ||
       (promptTokens as number) < 0 ||
       (promptTokens as number) > admission.maxInputTokens ||
-      (completionTokens as number) > admission.maxOutputTokens ||
-      (promptTokens as number) + (completionTokens as number) !== totalTokens
+      (completionTokens as number) > admission.maxOutputTokens
     ) {
       throw new Inf0OutcomeUnknownError();
     }
+    const totalTokens = (promptTokens as number) + (completionTokens as number);
 
-    const finish = result.choices?.[0]?.finish_reason;
+    const finish = result.done_reason;
     return {
       status:
         finish === 'stop' && (completionTokens as number) > 0 && content.trim()