Flush sinks and run periodic tasks when manual backend polling drains the queues - #989

ManualBackendWorker::poll() loops while (!_check_frontend_queues_and_cached_transit_events_empty()) poll_one(); and returns the moment the queues drain (include/quill/backend/ManualBackendWorker.h:161). poll(timeout) has the same shape. Each of those poll_one() calls therefore only ever takes the branch of BackendWorker::_poll() that processes a cached transit event.

Everything else the backend does lives in the branch a poll takes when it finds no cached transit events (include/quill/backend/BackendWorker.h:546-592): _flush_and_run_active_sinks(true, sink_min_flush_interval, SinkFlushReason::Periodic), _check_failure_counter(), _resync_rdtsc_clock() and, once the queues are also empty, _cleanup_invalidated_thread_contexts(), _cleanup_invalidated_loggers() and _try_shrink_empty_transit_event_buffers(). A polling loop built out of poll() can never reach it, because the loop exits exactly when a poll would have found nothing left to process.

So driving the backend the way the Basic Usage section of docs/manual_backend_worker.rst describes, "Periodically call poll() or poll(timeout)", as examples/manual_backend_worker.cpp does, never calls Sink::flush_sink() or Sink::run_periodic_tasks() while the application runs. A FileSink's records sit in the stdio buffer until it fills or shutdown() runs, so tail -f shows nothing and a crash loses the tail of the log. Dropped message notifications, the rdtsc resync BackendTscClock depends on, and the memory of removed loggers and retired thread contexts are skipped for the same reason. Calling poll_one() in your own loop is not affected: it hits the idle branch whenever there is nothing to process.

The fix polls once more after the loop exits because the queues drained. poll(timeout) returns instead when it runs out of budget with records still pending, so its incremental drain is unchanged and TEST_ManualBackendWorkerTimeoutPoll still sees a partial drain.

The regression test logs through a sink that counts flush_sink() and run_periodic_tasks() calls, then requires both counts to be non-zero after poll() and again after a poll(timeout) on already drained queues. It sets sink_min_flush_interval to 0 and log_timestamp_ordering_grace_period to 0 so the assertions do not depend on timing.

Verified on macOS arm64 with Apple clang 17, cmake -S . -B build -DQUILL_BUILD_TESTS=ON -DCMAKE_BUILD_TYPE=Debug then cmake --build build --target <target> -j 4. On master the new test fails at the first flush assertion with REQUIRE_GT( 0, 0 ); with the change TEST_ManualBackendWorkerPollSinkFlush reports 1 test case and 4 assertions passing. TEST_ManualBackendWorker (18 assertions), TEST_ManualBackendWorkerTimeoutPoll (3), TEST_ManualBackendWorkerThreadMigration (14) and TEST_ManualBackendWorkerErrorNotifier (2) all still pass.