Open questions

Still undecided. Resolve, then move to Locked decisions.

1. LightRAG storage / graph backend — unified Postgres (pgvector + AGE, self-host Hetzner) vs Neo4j + dedicated vector DB (Qdrant/Milvus). Depends on #2.
2. Scale shape — are "millions of files" aggregate across many small bases (→ unified PG ok) or millions in a single base (→ Neo4j + Qdrant)?
3. Object storage — Hetzner Object Storage vs AWS S3 vs Cloudflare R2 (security parity given presigned URLs; differentiators = compliance vs residency/latency vs egress).
4. Postgres host — self-host on Hetzner (required for AGE) vs Neon (serverless, only if graph goes to Neo4j).
5. ✅ Email — RESOLVED: Resend for MVP → AWS SES at scale (swappable behind notification/email).
6. Billing provider — PayPal only vs add Stripe (Stripe fits usage-metered billing).
7. Cognito user migration — existing users have no local password hash; seed via reset-password on cutover (or import).
8. Rename runtime/worker slice (the manager) to avoid clashing with the worker app.
9. Benchmark LightRAG at target scale before committing.

(The former #7, Gateway / channels, was answered by AGNT2-222 and moved to Locked decisions — see Channels.)