fix: bound uncached remainder by prompt-max(cached,write) and forward compact prompt_cache_key

dto/openai_response.go

@@ -270,12 +270,17 @@ type InputTokenDetails struct {
 // which field the upstream reported it in: Claude-derived conversions populate
 // CachedCreationTokens while OpenAI reports cache_write_tokens natively. Both
 // are billed at the cache-creation price; when both are present the larger
-// value wins so the same tokens are never double-counted.
+// value wins so the same tokens are never double-counted. Negative upstream
+// values are clamped to zero so they can never lower a charge.
 func (d InputTokenDetails) CacheCreationTokensTotal() int {
-	if d.CacheWriteTokens > d.CachedCreationTokens {
-		return d.CacheWriteTokens
+	total := d.CachedCreationTokens
+	if d.CacheWriteTokens > total {
+		total = d.CacheWriteTokens
 	}
-	return d.CachedCreationTokens
+	if total < 0 {
+		return 0
+	}
+	return total
 }
 
 type OutputTokenDetails struct {