fix(usage): OpenAI 缺失 input_tokens 时从 total 回推

- 同步与流式两条路径均在 input_tokens 为 0 且 total > output 时用 total - output 兜底
- 补充对应单元测试

crates/aether-ai-pipeline/src/finalize/standard/openai/stream.rs

@@ -1822,6 +1822,22 @@ mod tests {
         sequence_numbers
     }
 
+    #[test]
+    fn openai_usage_derives_missing_input_tokens_from_total() {
+        let usage = canonical_usage_from_openai_usage(Some(&json!({
+            "output_tokens": 177,
+            "total_tokens": 20_612,
+            "input_tokens_details": {
+                "cached_tokens": 19_840,
+            },
+        })))
+        .expect("usage should parse");
+
+        assert_eq!(usage.input_tokens, 20_435);
+        assert_eq!(usage.output_tokens, 177);
+        assert_eq!(usage.cache_read_tokens, 19_840);
+    }
+
     #[test]
     fn openai_cli_client_emitter_emits_doc_like_text_events() {
         let mut emitter = OpenAICliClientEmitter::default();

crates/aether-ai-pipeline/src/finalize/standard/stream_core/common.rs

@@ -89,7 +89,7 @@ pub fn resolve_identity(
 
 pub fn canonical_usage_from_openai_usage(value: Option<&Value>) -> Option<CanonicalUsage> {
     let usage = value?.as_object()?;
-    let input_tokens = usage
+    let mut input_tokens = usage
         .get("input_tokens")
         .or_else(|| usage.get("prompt_tokens"))
         .and_then(Value::as_u64)
@@ -129,6 +129,9 @@ pub fn canonical_usage_from_openai_usage(value: Option<&Value>) -> Option<Canoni
             .saturating_add(cache_creation_tokens)
             .saturating_add(cache_read_tokens),
     );
+    if input_tokens == 0 && total_tokens > output_tokens {
+        input_tokens = total_tokens.saturating_sub(output_tokens);
+    }
     Some(CanonicalUsage {
         input_tokens,
         output_tokens,