School attendance CDC includes soft-deletes for withdrawn students. How should gold enrollment counts stay accurate?

Sequencing CDC events by processing time in gold and deduplicating withdrawals using watermark filtering is unreliable, risks dropping late-arriving withdrawals during the watermark window, and defeats the streaming purpose of CDC. Soft-delete or tombstone markers mean the student is no longer active; gold metrics that power attendance dashboards and funding-related headcounts should reflect that withdrawal instead of treating every historical insert as still enrolled. Applying CDC deletes (or equivalent effective-end dating) keeps civic enrollment numbers accurate as students leave the district. Merging inserts and updates on student key into gold treating soft-deletes as insert-only rows keeps withdrawn students in active gold counts as if still enrolled, inflating current-day enrollment instead of tombstoning withdrawals. Sequencing CDC events by processing time in gold and deduplicating withdrawals using watermark filtering is operationally destructive to count integrity because late withdrawals can be evicted by the watermark, throwing away valid withdrawal signals along with stale events. Applying CDC soft-deletes so gold counts version withdrawn students with Type 2 effective-date history keeps withdrawn students in active enrollment totals instead of tombstoning them for current counts, so current-day headcounts stay inflated rather than reflecting who is actually enrolled. Exam caveat: CDC soft-deletes are first-class change events. Medallion gold for enrollment is not append-only history of every insert; it is a business-facing count of currently enrolled students (or clearly dated enrollment spans). Distinguish immutable bronze change logs from curated gold that must honor deletes. Operational check: Process a CDC batch that soft-deletes a known student_id, then recompute gold enrollment and confirm that student no longer appears in active counts while remaining students are unchanged. Verify tombstone or end-date fields are set as designed. Re-run with an insert-only batch and confirm counts rise only for true new enrollments.