Closes #1621

Summary

• Preserve scheduler rejection outputs when decode responses are produced in the same step().
• Add a regression test covering a prefill failure output followed by decode output in the same scheduler step.

Why

When prefill failed, _schedule_waiting() could return a terminal error RequestOutput for the failed request. If another request produced decode output in the same Scheduler.step(), the decode path replaced output.outputs instead of appending to it, dropping the prefill error output.

That left the failed request removed from scheduler state but without a finished output delivered to engine_core, so the client could wait indefinitely with no tokens and no error.

Changes

• Changed Scheduler.step() to append decode outputs with output.outputs.extend(outputs).
• Changed finished request ID handling to merge with output.finished_request_ids.update(finished_ids).
• Added TestSchedulerStepOutputs.test_decode_outputs_preserve_prefill_rejections.

Tests

.venv/bin/python -m pytest tests/test_scheduler.py -q

Result:

105 passed

Also verified the new regression test fails when the scheduler fix is reverted.