4.8.14 Internal Network Workers

thread.ss is an implementation-only module. Neither the public network interface nor its facade exports its helpers, marker or result wrapper.

spawn-network-thread(name, thunk) starts a named actor with a non-inherited thread-local marker. The marker remains set during cleanup. Application children start with independent locals; Network.close alone uses the marker to reject worker-side blocking shutdown before taking locks or changing state.

An outer with-catch aborts out of thunk before its handler runs. Consequently, dynamic finalizers finish before the handler logs Network worker failed at debug level through the existing /ensemble/network logger and returns a final NetworkThreadError containing the original exception : :t, including raised #f. No generic symbol represents failure. Normal results, including #f or returned exception objects, are unchanged. The internal wrapper is reserved for completed worker failures, not ordinary application data.

This prevents handled worker failures from reaching the actor’s unhandled-error stack-trace wrapper. It does not remove the runtime’s abortive root handler, change existing contextual logs inside workers, or introduce asynchronous cancellation. Owners still cancel cooperatively, close sockets to unblock IO, finish callbacks, and join before releasing borrowed resources.

network-thread-join!(thread, [timeout, timeout-value]) delegates to thread-join!/error and raises the exact exception from a completed wrapper. It also accepts ordinary application threads: normal results and native worker exceptions keep the shared join helper’s semantics. Omitted, false, relative and absolute timeouts behave as before. A private timeout sentinel preserves every explicit timeout value, even a NetworkThreadError supplied as that value. A worker-raised Timeout or native join-timeout exception is not translated into a timeout of the caller’s join.

Raw native joins now see a normal completion containing NetworkThreadError on worker failure. The two connection aggregate barriers deliberately inspect this result themselves: only a completed result or native uncaught-exception wrapper proves termination. They retain the first failure while joining remaining workers, without swallowing a join-operation failure or confusing a worker’s Timeout with an incomplete join. connection-shutdown! still suppresses only the exact stored parent-close failure, retaining distinct cleanup failures including #f. Tests that inspect thread state must not equate normal termination with success.

thread-test.ss covers ordinary results and returned exceptions, original raised object identity and #f, dynamic cleanup before completion, marker retention, relative/absolute/polling/omitted timeout semantics and worker-raised native timeout identity. Logger behavior is source-reviewed; the test does not mutate the shared global logging configuration or assert asynchronous sink delivery.

After the main process rebuilds stdlib, run:

./build.sh test std/ensemble/network/thread-test.ss
./build.sh test std/ensemble/network/...