fix: cascade correlation — fallback to _latest MITM usage

When the MITM can't extract a cascade ID from the intercepted request
(Content-Length: 0 / chunked encoding), usage is stored under '_latest'.
Now usage_from_poll and completions try the exact cascade_id first,
then fall back to '_latest' so MITM-captured tokens are actually used.

src/api/responses.rs

@@ -277,7 +277,12 @@ async fn usage_from_poll(
     output_text: &str,
 ) -> Usage {
     // Priority 1: MITM intercepted data (most accurate — includes cache tokens)
-    if let Some(mitm_usage) = mitm_store.take_usage(cascade_id).await {
+    // Try exact cascade_id match first, then fall back to "_latest" (unmatched)
+    let mitm_usage = match mitm_store.take_usage(cascade_id).await {
+        Some(u) => Some(u),
+        None => mitm_store.take_usage("_latest").await,
+    };
+    if let Some(mitm_usage) = mitm_usage {
         tracing::debug!(
             input = mitm_usage.input_tokens,
             output = mitm_usage.output_tokens,