fix: keep the status of synchronous responses on late client failures - #1608

Refs #1601.

Cause

The FINE logs from #1602 captured a failure in CI. In SecurityTestCase, the client gets the 403 head and the first bytes of the server's HTML error page, so resource.get() returns. The test then calls resource.release(), which stops the ClientResource's own client while the rest of the body is still arriving. Closing the connection fails the in-flight call. Since the head was already received, dispatchResponseFailureCallback() then set the status to CONNECTOR_ERROR_COMMUNICATION (1001), even for synchronous requests. That's the same Response object the test reads afterwards, so its 403 turned into 1001.

.708  Received the response head 403 Forbidden
.708  Stopping the Netty client          (test calls release())
.708  Closed the connection
.709  The exchange failed: IOException "Connection closed by peer"

The passing runs show the same order, but there the end of the response arrives before the close.

Changes

• NettyClientCall:
  • Synchronous requests: a failure after the head no longer changes the status. It reaches the response entity stream only, as an IOException on read.
  • Asynchronous requests: unchanged. The failure still sets 1001 and invokes the onResponse callback.
  • Empty bodies: after the head of a response with Content-Length: 0 (and no chunked encoding), a failure on the connection isn't reported as a response failure.
  • Completion order: when the response completes, the connection is released before the body stream ends. A synchronous caller resumes when the body ends, so its next request now always finds the connection back in the pool. Before, it could open a new connection, which made the empty-body test time out on CI (fixed in 62e72d6, after the rebase on 3.0).
• NettyClientHandler and NettyClientHelper: stop() marks its connections before closing them. A call cut short by the stop now fails with IOException("Client stopped") instead of "Connection closed by peer". For HTTP/2 and HTTP/3 streams, the mark on the parent connection is checked too.
• ClientLateFailureTestCase (new), with a raw socket server:
  • The server sends a 403 head and a first chunk, waits until the client stops, then sends the last chunk. The test checks that the 403 is kept, and that the entity stream fails with "Client stopped". Without the fix it fails 3 times out of 3, with the CI symptom: expected: <Forbidden (403)> but was: <Communication Error (1001)>.
  • Two Content-Length: 0 responses on one connection: both 403 with an empty entity, and the connection is reused.
• changes.md: entry under Bugs fixed.

Http2CancellationTestCase still passes: it's asynchronous, so its callback and status still get 1001.

Validation

./mvnw -B -pl org.restlet -P default,test verify: 2217 tests pass, the coverage check passes, and SpotBugs still reports 8 findings.

🤖 Generated with Claude Code