fix(sse): include prompt cache usage fields on message_stop fallback path (#10545)

Mirrors the existing prompt_tokens_details cache-field mapping from the message_delta finish path into the message_stop fallback, so OpenAI-compatible clients see cache_read/cache_creation counters when the finish signal arrives without usage on the same event. Fixes #10535.

Validated in an isolated worktree boarded onto origin/release/v3.8.50 (0 conflicts, 2 files):
- 16/16 focused tests pass (translator-resp-claude-to-openai.test.ts), including the new regression test for this exact fallback path.
- check-file-size, check-changelog-integrity: OK.
- typecheck:core: clean.
- check-complexity / check-cognitive-complexity: OK, both under baseline.

Co-authored-by: NahuSaruf <NahuSaruf@users.noreply.github.com>

open-sse/translator/response/claude-to-openai.ts

@@ -332,6 +332,8 @@ export function claudeToOpenAIResponse(chunk, state) {
       if (!state.finishReasonSent) {
         const finishReason =
           state.finishReason || (state.toolCalls?.size > 0 ? "tool_calls" : "stop");
+        const cachedTokens = state.usage?.cache_read_input_tokens || 0;
+        const cacheCreationTokens = state.usage?.cache_creation_input_tokens || 0;
         const usageObj =
           state.usage && typeof state.usage === "object"
             ? {
@@ -347,6 +349,16 @@ export function claudeToOpenAIResponse(chunk, state) {
                         },
                       }
                     : {}),
+                  ...(cachedTokens > 0 || cacheCreationTokens > 0
+                    ? {
+                        prompt_tokens_details: {
+                          ...(cachedTokens > 0 ? { cached_tokens: cachedTokens } : {}),
+                          ...(cacheCreationTokens > 0
+                            ? { cache_creation_tokens: cacheCreationTokens }
+                            : {}),
+                        },
+                      }
+                    : {}),
                 },
               }
             : {};

tests/unit/translator-resp-claude-to-openai.test.ts

@@ -403,6 +403,40 @@ test("Claude stream: message_stop falls back to tool_calls when tool use already
   assert.equal(result[0].choices[0].finish_reason, "tool_calls");
 });
 
+test("Claude stream: message_stop includes prompt_tokens_details when usage arrived on an earlier message_delta without stop_reason (#10535)", () => {
+  const state = createState();
+  claudeToOpenAIResponse(
+    { type: "message_start", message: { id: "msg1", model: "claude-sonnet-4-6" } },
+    state
+  );
+
+  // Usage lands on a message_delta that carries no stop_reason (e.g. an
+  // upstream that reports usage and the finish signal in separate events),
+  // so the finalChunk branch in the message_delta case never runs and
+  // finishReasonSent stays false.
+  const deltaResult = claudeToOpenAIResponse(
+    {
+      type: "message_delta",
+      delta: {},
+      usage: {
+        input_tokens: 8,
+        output_tokens: 5,
+        cache_read_input_tokens: 2000,
+        cache_creation_input_tokens: 0,
+      },
+    },
+    state
+  );
+  assert.equal(deltaResult, null);
+
+  const result = claudeToOpenAIResponse({ type: "message_stop" }, state);
+
+  assert.equal(result[0].usage.prompt_tokens, 2008);
+  assert.equal(result[0].usage.completion_tokens, 5);
+  assert.equal(result[0].usage.total_tokens, 2013);
+  assert.equal(result[0].usage.prompt_tokens_details.cached_tokens, 2000);
+});
+
 test("Claude stream: unsupported events return null", () => {
   assert.equal(claudeToOpenAIResponse({ type: "error" }, createState()), null);
 });