Migrate storage layer from Snowflake/DynamoDB/SQS to Postgres/NATS (graph backend fixes)

[jonathanhaaswriter]

Summary

• start from feat/postgres-nats-migration and keep Postgres scoped to storage and warehouse responsibilities instead of treating it as a graph backend
• move read-only SQL validation into backend-agnostic warehouse helpers and update API, agent, and CLI call sites accordingly
• wire in the pgx stdlib driver plus vendored dependencies required for Postgres runtime support, and refresh generated config env docs

Testing

• CEREBRO_DEVEX_BASE_REF=writer/main python3 ./scripts/devex.py run --mode changed --base-ref writer/main

[jonathanhaaswriter]

I found two blocking issues in the current head:

• internal/app/app_init_core.go:initPostgres still couples DATABASE_URL to a.Warehouse, so storage DB bootstrap overrides WAREHOUSE_BACKEND and prevents the configured warehouse from being initialized. initWarehouse also still lacks a snowflake branch.
• internal/jobs/worker.go switched retry backoff back to ExtendVisibility(delay), but internal/jobs/queue_nats.go:ExtendVisibility ignores the delay argument and only calls Msg.InProgress(). Retryable failures therefore won't honor exponential backoff under JetStream.

I think both need to be resolved before this is ready.

[jonathanhaaswriter]

I think we still have two problems here:

• effectivePostgresURL() prefers JOB_DATABASE_URL over DATABASE_URL, so if both are set the app repos silently connect to the jobs DB instead of the normal app DB
• the local findings fallback changed from SQLite to in-memory, which drops the persisted local-mode behavior we had before

Can we keep app storage on DATABASE_URL and preserve the SQLite fallback when Postgres isn't configured?