fix(ai): run gateway tool calls whatever finish reason comes beside them

The Coach loop runs a round's tool calls only when the round finished with
"tool_calls". The Anthropic and Codex clients already derive that from the
calls themselves; the OpenAI-compatible client copied the server's
finish_reason, and vLLM, Ollama and LiteLLM can report "stop" beside a
parsed tool call. The loop then took the round as the answer, the call never
ran, and what the person saw was whatever text came with it.

A reply that carries tool calls now reports "tool_calls", as on the other
wires.

Refs #1276

src/lib/ai/coach/__tests__/tools-loop-always-reasoning-gateway.test.ts

@@ -312,6 +312,13 @@ describe("#1276: a gateway model that always reasons", () => {
     },
   );
 
+  it("runs a tool call the server reports with finish_reason stop", async () => {
+    toolFinish = "stop";
+    const out = await turn("medium");
+    expect(executeCoachTool).toHaveBeenCalledTimes(1);
+    expect(out.result.content).toBe(ANSWER);
+  });
+
   it("fails as reasoning_only, never with the thought, when no answer ever comes", async () => {
     parser = "raw";
     reasoningWords = 20_000;

src/lib/ai/openai-client.ts

@@ -554,7 +554,12 @@ export class OpenAIClient implements AIProvider {
       model: this.config.model,
       providerType: this.type,
       ...(toolCalls ? { toolCalls } : {}),
-      finishReason: mapFinishReason(choice?.finish_reason),
+      // #1276 — the tool calls themselves decide, as on the Anthropic and
+      // Codex wires: vLLM, Ollama and LiteLLM can put `stop` beside them,
+      // and the loop runs calls only on `tool_calls`.
+      finishReason: toolCalls
+        ? "tool_calls"
+        : mapFinishReason(choice?.finish_reason),
       ...(params.reasoning
         ? {
             reasoning: {