Preserve active solves across deploys

[coder13]

Summary

• preserve authenticated room membership through brief socket restarts, including the current scramble, waiting/competing state, and room admin
• keep an in-progress browser timer mounted while the room socket reconnects
• save completed solves in a local outbox until the client rejoins the room and receives a persistence acknowledgement
• make result submission idempotent with stable submission IDs and immutable attempt keys, including recovery when room advancement fails after the result is saved
• disable the legacy Grand Prix mode unless GRAND_PRIX_ENABLED=true is explicitly configured
• expose dependency-aware API and socket health endpoints; optional PostgreSQL failure reports degraded without taking the primary MongoDB/Redis path out of service
• classify deploys by changed paths, roll socket before API/static when both change, and skip application replacement for documentation-only changes
• wait for backing-service health before migrations and restore exact pre-deploy images with the previous Compose configuration if rollout readiness fails

User-visible behavior

Socket deploys still cause a brief transport reconnect, but they no longer end the user's room session. A solve already in progress continues timing locally. If it finishes while disconnected, the next solve is blocked and the completed time is retained on the device until it can be submitted safely.

The local result is marked as delivery-attempted before its first emit, so an uncertain submission cannot be discarded accidentally. A definitive server rejection remains discardable across refreshes. Invalid manual times stay in the input for correction and never enter the outbox.

Grand Prix rooms are hidden and rejected in production while the mode is disabled. Normal rooms are unaffected.

Deployment behavior

DEPLOY_TARGET=auto resolves documentation-only changes to none, browser-only changes to api, and shared/server/deployment changes to all. Missing runtime containers promote an automatic deploy to a full bootstrap. Full deploys replace socket before API/static, health-check each service, and gracefully reload nginx after each handoff.

Rollback tags preserve the exact image IDs that were running before deployment. Rollback also uses Compose files from the pre-deploy commit, so a bad command, environment definition, or health check can be reverted along with the image. Database migrations are intentionally not reversed and must remain backward-compatible.

Validation

• manually restarted the API and socket processes multiple times during active solves; timing continued and completed results submitted after reconnect
• yarn lint
• yarn test — 55 client tests and 70 server tests
• yarn build
• Prisma schema validation, migration deploy, and drift check against local PostgreSQL
• scripts/test-deploy-classifier.sh
• scripts/test-deploy.sh
• Bash syntax checks for all deploy scripts
• production and development Compose parsing with Docker's default context
• yarn cypress:run — 6 tests across both specs
• git diff --check