[Jens-G]

THRIFT-6443

dispatch_rejection/too_many_pending_tasks_does_not_end_the_server in TServerIntegrationTest fails intermittently on Windows with "no client was dropped; the pool was not driven into rejection". Since 2026-10-19, 15 of the 45 AppVeyor jobs on commits that contain the case failed that way (MSVC2022 and MinGW), and none of the 15 jobs without it. The GitHub Actions MSVC build is hit too, for example on master 47fe9968f and d01c127d0.

The server does what it should; the test misses the drops. With a non-blocking dispatch, serve() rejects and closes a client right after accepting it, while the test client still writes its request in several send() calls. When the close gets there first, the send fails ("An established connection was aborted by the software in your host machine" in every failing log), the catch (...) discards that client, and it never reaches the end-of-stream check. When that happens to every rejected client, the check finds no drop.

The test now counts a fired client whose send fails with a transport error other than a timeout as dropped, like one that reads end-of-stream or a reset. Test-only change.

Checked in thrift:jammy (Release build of TServerIntegrationTest):

variant	result
master test, interleaving forced (each fired client writes one byte, waits 200 ms, then sends)	5 of 5 fail with the CI message; the sends fail with "Broken pipe"
this branch, same forced interleaving	5 of 5 pass
master test, unchanged	dispatch_rejection 15 of 15 pass
this branch, unchanged	dispatch_rejection 15 of 15 pass, whole module "No errors detected"

The forced interleaving was a local experiment only and is not part of the change. On Windows, only CI can show the effect: the MSVC job and AppVeyor of this PR.

• Apache Jira ticket: THRIFT-6443
• Single commit
• Code change, so the commit carries no CI-skip marker.

🤖 Generated with Claude Code