test(kiro): cover the real upstream shape that omits tokenUsage

CodeWhisperer does not report per-turn cache detail. Proven statically
without credentials: parseTokenUsage reads totalTokens as required, so a
parsed authoritative frame always yields usage.totalTokens, and no recent
Kiro usage row carries it (or any cache field) in a 5691-row live sample.

The 4736 rows that DO carry a large usage.totalTokens all fall in a
06-30..07-01 window, from when 7374c3a wrote contextUsagePercentage x
contextWindow straight into totalTokens. That figure later moved to the
dedicated contextTotalTokens field, which persistence dropped — which is
why the large numbers disappeared from Logs.

This test pins the live shape: contextUsagePercentage present, tokenUsage
absent. The turn stays estimated, invents no cache split, and still
reports a contextTotalTokens checkpoint larger than the per-turn total.
The cache read/write parser stays in place, so if upstream ever starts
reporting the split it surfaces with no further code change.

Author: lidge-jun

tests/kiro-stream.test.ts

@@ -1242,6 +1242,27 @@ describe("kiro adapter — parseStream", () => {
     expect(done.estimated).toBe(true);
   });
 
+  test("a real-shaped Kiro turn without tokenUsage still reports a cumulative context checkpoint", async () => {
+    // This is the shape live CodeWhisperer actually sends: contextUsagePercentage but NO
+    // tokenUsage (proven statically — parseTokenUsage reads totalTokens as required, and no
+    // recent kiro usage row carries usage.totalTokens or any cache field). The per-turn
+    // numbers therefore stay small estimates, and contextTotalTokens is the ONLY signal of
+    // real context occupancy. It must be present so Logs can show cumulative growth.
+    const adapter = createKiroAdapter(provider);
+    await adapter.buildRequest(parsedWith([{ role: "user", content: "x".repeat(4_000) }]));
+    const done = await doneUsage(
+      adapter,
+      eventFrame({ content: "answer" }),
+      eventFrame({ contextUsagePercentage: 42 }, "metadataEvent"),
+    );
+    expect(done.estimated).toBe(true);
+    // No fabricated cache detail when upstream reports none.
+    expect("cacheReadInputTokens" in done).toBe(false);
+    expect("cacheCreationInputTokens" in done).toBe(false);
+    // The checkpoint exceeds the small per-turn total, which is the whole point.
+    expect(done.contextTotalTokens).toBeGreaterThan(done.inputTokens + done.outputTokens);
+  });
+
   test("authoritative metadata token usage overrides estimates and preserves cache splits", async () => {
     const adapter = createKiroAdapter(provider);
     await adapter.buildRequest(parsedWith([{ role: "user", content: "x".repeat(700) }]));