fix(thinking): send explicit thinking:{type:adaptive} alongside output_config.effort

tests/translator/__snapshots__/golden-request.test.js.snap

@@ -118,6 +118,9 @@ exports[`GOLDEN request: OpenAI → Claude > reasoning_effort → adaptive outpu
       "type": "text",
     },
   ],
+  "thinking": {
+    "type": "adaptive",
+  },
 }
 `;
 

tests/translator/thinking-unified.test.js

@@ -55,10 +55,14 @@ describe("extractThinking", () => {
 });
 
 describe("applyThinking per provider format", () => {
-  it("claude 4.6+ → adaptive output_config (no budget_tokens)", () => {
+  it("claude 4.6+ → adaptive thinking + output_config (no budget_tokens)", () => {
     const out = apply("claude", "claude-opus-4.7", { reasoning_effort: "high" }, "claude");
     expect(out.output_config).toEqual({ effort: "high" });
-    expect(out.thinking).toBeUndefined();
+    // Anthropic: on Opus 4.6/4.7/4.8 and Sonnet 4.6 thinking stays OFF unless
+    // thinking:{type:"adaptive"} is sent explicitly; output_config alone is not
+    // enough (and Anthropic-compatible shims like Copilot default off even on
+    // Sonnet 5). Both fields together are the documented adaptive shape.
+    expect(out.thinking).toEqual({ type: "adaptive" });
   });
   it("claude haiku → enabled+budget", () => {
     const out = apply("claude", "claude-haiku-4.5", { reasoning_effort: "high" }, "claude");