Filed 2026-10-10 (golang/go#82120). Open/unfixed. When http2.Transport.RoundTrip returns an error after response DATA was already buffered into the stream pipe (classic: short deadline + preempted caller, or cancel during cleanup), connection-level inflow taken in processData is never refunded β cleanupWriteRequest uses CloseWithError (leaves unread bytes) instead of BreakWithError+WINDOW_UPDATE as Body.Close does.
Concrete from the primary:
- Repro on x/net v0.60.0 (and v0.37.0): cancel path leaks 16 KiB per errored RoundTrip;
-collapseexhausts the ~1 GiB default conn window until peerRST_STREAM(FLOW_CONTROL_ERROR). - Production symptom (h2c client vs bRPC, ~100 ms deadlines): after ~a week, ~2 KB
remote_conn_window_left, then near-total request failure on that connection without closing it. - Existing #49645 tie-break covers
cs.abortvsrespHeaderRecv, notctx.Done/reqCancelvsrespHeaderRecv. - Fix sketch: abandon buffered response the way
transportResponseBody.Closedoes (BreakWithError + refund) on every nil-Response error path after stream create.