• fix exit-node v1.9.4: مدارا با TLS ungraceful close (peer closed without close_notify) از سمت host exit-node ([#585](https://github.com/therealaleph/MasterHttpRelayVPN-RUST/issues/585) از @gregtheph): در v1.9.4، کاربری که exit node رو با درست‌ترین config setup کرد، در log می‌دید `WARN exit node failed for https://chatgpt.com/: io: peer closed connection without sending TLS close_notify — falling back to direct Apps Script` + سپس fallback به Apps Script که خود نمی‌تونه ChatGPT رو reach کنه، در نتیجه decoy/no-json error. علت: rustls سختگیر است درباره‌ی TLS shutdown — وقتی peer (the exit node) underlying TCP رو می‌بنده بدون اول send کردن TLS close_notify alert، rustls `io::ErrorKind::UnexpectedEof` می‌فرسته. کد ما در `read_http_response` این error رو propagate می‌کرد به‌عنوان hard error. حالا UnexpectedEof به‌صورت graceful EOF (مشابه `n == 0`) درمان می‌شه — اگر body completed شده با Content-Length، response درست برمی‌گرده. اگر mid-body close بود، error real (truncation) همچنان propagate می‌شه. ۴ regression test جدید (شامل UnexpectedEof tolerance + envelope unwrap exit_node). 173 lib tests + 33 tunnel-node tests pass. --- • Fix v1.9.4 exit-node: tolerate ungraceful TLS close (peer closed without close_notify) on the exit-node path ([#585](https://github.com/therealaleph/MasterHttpRelayVPN-RUST/issues/585) by @gregtheph): in v1.9.4, users with a correctly-configured exit-node deployment saw `WARN exit node failed for https://chatgpt.com/: io: peer closed connection without sending TLS close_notify — falling back to direct Apps Script` in the log, followed by a fallback to direct Apps Script which can't reach ChatGPT either, resulting in the decoy/no-json error. Root cause: rustls is strict about TLS shutdown — when the peer (the exit-node's host) closes the underlying TCP without first sending a TLS close_notify alert, rustls surfaces this as `io::ErrorKind::UnexpectedEof`. Our code in `read_http_response` was propagating this as a hard error rather than treating it as graceful EOF. Now `UnexpectedEof` is handled like `n == 0`: if the body has been fully received per Content-Length, the response returns successfully; if it's a real mid-body truncation, the error still propagates as `BadResponse`. Same handling added to the chunked reader and the no-framing reader. Four regression tests cover the new behavior (UnexpectedEof tolerance for Content-Length and no-framing branches + exit-node envelope unwrap success and error paths). 173 lib tests + 33 tunnel-node tests passing.