fix: prevent SSE stream corruption by requesting identity encoding

Add 'accept-encoding: identity' to upstream request headers to prevent
Cloudflare/Anthropic CDN from returning gzip-compressed SSE responses.

When the upstream returns gzip data without a Content-Encoding header
(observed with some CDN edge configurations), the stream handler's
chunk.toString() call decodes raw gzip bytes as UTF-8, replacing any
byte >= 0x80 that isn't valid UTF-8 with U+FFFD (ef bf bd). This
silently corrupts the entire SSE stream.

Explicitly requesting identity encoding is the simplest and most
reliable fix — it prevents compression at the source rather than
trying to detect and decompress it after the fact.

Closes #1030

src/services/relay/claudeRelayService.js

@@ -1494,11 +1494,16 @@ class ClaudeRelayService {
     const contentLength = Buffer.byteLength(bodyString, 'utf8')
 
     // 构建最终请求头（包含认证、版本、User-Agent、Beta 等）
+    // Force identity encoding to prevent upstream (Cloudflare) from returning
+    // gzip-compressed responses without a Content-Encoding header, which causes
+    // binary data to be silently corrupted by UTF-8 text decoding in the stream
+    // handler. See: https://github.com/Wei-Shaw/claude-relay-service/issues/1030
     const headers = {
       host: 'api.anthropic.com',
       connection: 'keep-alive',
       'content-type': 'application/json',
       'content-length': String(contentLength),
+      'accept-encoding': 'identity',
       authorization: `Bearer ${accessToken}`,
       'anthropic-version': this.apiVersion,
       ...finalHeaders