(feat): Integration And Push

[Koded0214h]

☐ VPS deployment: Docker Compose on Contabo, env vars, HTTPS ☐ Smoke test: create merchant → plan → subscribe → bill → charge → webhook → gym app shows active ☐ Recovery test: force failure → transfer VA issued → simulate push → reconcile → active ☐ payment_uncertain test: timeout → freeze → verify → resolves ☐ Hard bug fixes from integration
☐ Proration end-to-end: upgrade mid-cycle → correct charge fires ☐ Two-merchant isolation proof: same instance, fully isolated data ☐ billing_lag_seconds metric: log time from next_bill_date to charge_attempt created ☐ queue_depth log: consumer lag on billing partition every 60s ☐ recovery_rate metric: past_due healed / total past_due ☐ Integration test: full billing cycle (create → bill → heal) ☐ Integration test: payment_uncertain → verify pass → active

(tests): All tests green, connected with the server bash scripts, integration tests and unit tests

NOTES

Deployment fixed

• HTTPS was never actually wired up despite a valid cert sitting on disk — nginx only had a port-80 block, and that block was proxying to the wrong app (an unrelated ship-cli project sharing the box). Fixed both; https://somba.ddns.net now correctly terminates TLS and routes to Somba.
• Docker port mismatch: PORT env var was used for both the host-side publish port and leaked into the container, so uvicorn bound to the wrong port internally — app was unreachable. Fixed.
• The scheduler wasn't running at all — only the API server was in docker-compose.yml. Nothing was ever billing, delivering webhooks, or reconciling in production. Added it as a proper service.
• Nomba LIVE credentials are live and verified working, including the parent/sub-account header split your hackathon email specified (auth uses the parent account, resource calls use the sub-account — the client mixed these up before).

Real bugs found and fixed (via actual deployment, not local tests)

1. Two competing billing-sweep implementations shared one lock; only the dead-end one was scheduled — subscriptions would never have been charged automatically.
2. Two schema-drift bugs: customers and webhook_deliveries tables were missing columns the ORM models expected (migrations never wrote them). Added a permanent regression test that diffs every migration against the live models.
3. Subscriptions never advanced their billing period after a successful charge — the most serious one. Every subscription would've been billed again the very next day, and the second charge would have crashed outright on a DB constraint. Fixed, with a regression test proving two consecutive billing cycles now work.
4. ChargeAttempt.order_reference collided on retries — any second attempt on the same intent (a retry, or reprocessing an "uncertain" charge) crashed with a UniqueViolation. Fixed.
5. VA (virtual account) creation was never actually wired up — the recovery engine's "transfer" path just emitted an event; nothing ever called Nomba to issue the account. Now it does, using the correct (undocumented-in-the-obvious-place) parent-account endpoint variant for sub-accounts — confirmed against live Nomba, real VA numbers get issued.
6. Fixed a bug in my own test script (reused a hardcoded fake transaction ID across runs, which tripped Somba's real duplicate-detection).

Webhook — tested as requested

Stood up a real echo receiver, drove a real delivery through the deployed emitter, and independently recomputed the HMAC-SHA256 signature against the merchant's secret — byte-for-byte match. Signing and
delivery are solid.

TESTS

To run it yourself

bash scripts/live/run_all.sh # full suite
bash scripts/live/02_recovery_test.sh # any script individually