test(reasoning): use temperature=0 to prevent Metal flap in ReasoningBudgetTest

The two tests that assert reasoning_content is present relied on the Qwen3
model sampling into a <think> block, which is non-deterministic at the
default temperature. Metal (macOS arm64) GPU arithmetic occasionally produces
different logit distributions than CPU backends and can sample a non-thinking
first token, causing a spurious failure.

Adding withTemperature(0.0f) (greedy sampling) makes both tests deterministic
across all platforms. The llama.cpp copy-loop bug (per-request
reasoning_budget_tokens ignored) is still present in b9739 — only the test
stability is improved here.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017xJt9CFqWYC4tWmC13YwVv

Author: claude

src/test/java/net/ladenthin/llama/ReasoningBudgetTest.java

@@ -97,11 +97,17 @@ public static void tearDown() {
      * Qwen3 enters thinking mode by default. With {@code reasoning_format=deepseek} set
      * at model level, the thinking tokens must appear in {@code reasoning_content} and
      * the final answer must appear in {@code content}.
+     *
+     * <p>{@code temperature=0} (greedy sampling) is used so the model deterministically
+     * enters the {@code <think>} block on every platform, including Metal (macOS arm64)
+     * where GPU floating-point arithmetic can produce slightly different logit
+     * distributions and occasionally sample a non-thinking first token.
      */
     @Test
     public void testThinkingDefault_reasoningContentAndAnswerPresent() {
         InferenceParameters params = new InferenceParameters("")
                 .withMessages(null, Collections.singletonList(new Pair<>("user", "What is 2+2?")))
+                .withTemperature(0.0f)
                 .withNPredict(N_PREDICT);
 
         String json = model.chatComplete(params);
@@ -132,11 +138,17 @@ public void testThinkingDefault_reasoningContentAndAnswerPresent() {
      * that is the signal to remove this test and enable
      * {@link #testReasoningBudgetZero_expectedBehavior_suppressesThinking}.
      * Tracked in <a href="https://github.com/ggml-org/llama.cpp/pull/23116">llama.cpp PR #23116</a>.
+     *
+     * <p>{@code temperature=0} (greedy sampling) is used so the model deterministically
+     * enters the {@code <think>} block on every platform. Without it, Metal (macOS arm64)
+     * occasionally samples a non-thinking first token even when the budget is unlimited
+     * (due to the bug), causing a spurious test failure.
      */
     @Test
     public void testReasoningBudgetZero_parameterAccepted_thinkingNotSuppressed() {
         InferenceParameters params = new InferenceParameters("")
                 .withMessages(null, Collections.singletonList(new Pair<>("user", "What is 2+2?")))
+                .withTemperature(0.0f)
                 .withReasoningBudgetTokens(0)
                 .withNPredict(N_PREDICT);