When a custom stdout or stderr splitter processes data or flushes asynchronously, PythonShell can emit close and invoke its end() callback before the final output event. A caller that reads its collected messages in the completion callback therefore receives incomplete output. The current order can be close → end callback → message("tail"); it should be message("tail") → close → end callback.

Track completion of each active splitter's readable side alongside the existing source-stream and process-exit checks. Keep waiting for the original streams as well, so a splitter that ends early does not truncate the stderr collected for process errors. No public API, dependency, or newline parsing changes.

Regression tests use real Python processes and a child-process close barrier, without sleeps or mocked process events. They cover stdout/stderr, asynchronous _transform/_flush, exit codes 0/7, and complete stderr collection when a splitter ends before its source.

Validation on macOS arm64, Node.js 22.23.1 and Python 3.9.6:

• Unmodified master 1cf1bfcac6ccde265bf0ac0e422fc54379941ed0: 43 existing tests pass.
• New asynchronous-output regressions before the fix: 8 fail with output arriving after completion.
• Focused suite after the fix: 9 pass.
• Full npm test: 52 pass.
• Prettier and git diff --check: pass; generated TypeScript declarations match master.

For the full suite, PYTHONPYCACHEPREFIX was set to a writable local directory for Python's syntax-check cache. Windows and Linux were not run locally.

Related work checked: #332 changes newline splitting; #329 flushes splitters in its timeout path. The eight asynchronous-output regressions still fail on both current PR heads without this fix. This PR addresses normal process completion, including nonzero exit codes.