The most trivial schema with a parent-child relationship either has foreign key constraints or update anomalies. There's no way to "sanely create" a normalized schema that doesn't have such relationships.

Performance impact may or may not matter. You don't discard valuable techniques and decades of experience for a "might happen." You measure and identify actual performance and adjust with facts. All indexes have both a performance benefit and a performance cost, incurred in the first case when selecting, and in the second case when updating. The performance benefit and cost of foreign keys comes from the index created. You get protection from update anomalies almost for free.