fix(agents): default size probe stays residual-free after claude row

The anthropic/claude-sonnet-4 fixture now resolves to the claude family,
so the default column silently measured the task_guidance block. Probe
openai/gpt-4.1 instead (default family until the CL-8310 gpt row lands)
and assert the default column carries no family residual.

Author: TheGreatAxios

src/agent/prompt-sizes.test.ts

@@ -343,9 +343,12 @@ describe("grok tool-budget residual (CL-8297)", () => {
   });
 
   test("default-family and orchestrator prompts carry no tool budget", () => {
-    expect(assembleDirectorPrompt("builder", "default")).not.toContain(
-      "Tool budget:",
-    );
+    const defaultPrompt = assembleDirectorPrompt("builder", "default");
+    expect(defaultPrompt).not.toContain("Tool budget:");
+    // The default probe (openai/gpt-4.1) resolves to the default family, so
+    // the default column carries no family residual at all — not the claude
+    // task_guidance block either.
+    expect(defaultPrompt).not.toContain("<task_guidance>");
     expect(assembleDirectorPrompt("skywalker", "grok")).not.toContain(
       "Tool budget:",
     );

src/agent/prompt-sizes.ts

@@ -57,9 +57,11 @@ export const CANONICAL_PROMPT_ENV: EnvironmentInfo = {
 };
 
 const GROK_PROVIDER = { providerName: "xai/default", model: "grok-4.6" };
+// Default-family probe: openai/gpt-4.1 resolves to the default family (the
+// gpt row lands later in CL-8310), so the default column carries no residual.
 const DEFAULT_PROVIDER = {
-  providerName: "anthropic",
-  model: "claude-sonnet-4",
+  providerName: "openai",
+  model: "gpt-4.1",
 };
 
 /** Families in the size table: default assembly vs Grok (+finish-bias note). */