Async deploy pipeline + port recycling + central config

[atharv10032]

Summary

Three related changes to make the deploy flow correct under real-world conditions:

• Async deploy. POST /projects/submit returns 202 {deploymentID, status: "pending"} immediately; the build runs in a detached goroutine with a bounded worker pool (8 concurrent) and panic recovery. Clients poll GET /projects/:id (auth-scoped) for state. Running the build inline was timing out browsers / proxies on real repos (docker build + Azure NSG poll runs into minutes).
• Port allocator overhaul.
  • Reuses db.GetCollection's pooled Mongo client (was opening + tearing down a fresh mongo.Connect/Disconnect per allocation).
  • Atomic claim via InsertOne against a unique index on ports.port (replaces the upsert-with-$setOnInsert that could silently let two concurrent claims both think they won).
  • Recycle pool: DELETE /projects/:containerName now actually frees the port via db.ReleasePort, and allocation tries the recycle pool (lowest available first) before the watermark scan.
  • Fixes the broken DeleteProjectByContainerName query: containername → container_name to match the Project struct's bson tag (the old form never matched, silently orphaning project docs).
• Central config. New internal/config package loads .env exactly once at startup via sync.Once. Removed the duplicate godotenv.Load calls from utils/jwt.go and utils/port_manager.go's init() (which ran before main.go's load and read empty env vars).

Breaking change for the client

• Before: POST /projects/submit blocked for minutes, returned 200 {url}.
• After: POST /projects/submit returns 202 {deploymentID, status: "pending"}. Client polls GET /projects/:id until status is succeeded (read hosted_url) or failed (read deploy_error).

autoship-client side is intentionally out of scope here; will follow in a separate PR.

Operator notes (one-time, before bringing this build up)

db.EnsurePortsIndex will fail at startup if the existing ports collection has duplicate port values (a remnant of the prior buggy upsert). Either drop the collection (loses existing reservations) or dedupe in place before deploying.

Test plan

• go build ./... and go vet ./... green locally
• Static deploy: submit, poll until succeeded, hit Blob URL
• Dynamic deploy: submit, poll through deploying → succeeded, hit hosted URL
• Concurrent submissions (≥9 at once): first 8 run, 9th queues — no 500s
• DELETE deployment: verify ports doc flips to status: "available" and Project doc is gone
• Second deploy reuses the freed port (recycle pool working)
• AWS regression with CLOUD_PROVIDER=aws

Known follow-ups (FUTURE comments in code)

Pointers for the next pass:

• services/dynamic.go near finalCmd.Run() — docker bind failure leaves a phantom port reservation; needs bound bool + defer db.ReleasePort.
• services/dynamic.go line ~227 — cloud.Get().AuthorizePort(hostPort) redundant after GetOrReserveValidFreePort already authorized internally (Azure NSG poll wastes ~10-30s).
• services/clone.go — clone path static/<user>/<repo> has no deploy id; concurrent same-repo deploys race on the same dir.
• api/auth.go GitHubCallback — userInfo["login"].(string) panics if login is missing/non-string.
• api/auth.go GitHubCallback — hardcoded http://localhost:3000/dashboard redirect; needs FRONTEND_URL env var for prod.
• services/dynamic.go Go branch of GenerateDockerfile — installCmd builds app then CMD runs whatever startCommand says; incoherent for Go projects.
• services/dynamic.go containerName — time.Now().Unix() is seconds-resolution; same-repo deploys landing in same second collide.
• api/auth.go Signup — no input validation; empty email/password create unusable user.

🤖 Generated with Claude Code

[Ashmit-Kumar]

LGTM. Next: Fix the auth module