fix(container): don't read PORT - the base image sets it to 34197 (#120)

[wormeyman]

Production is down, and this is the fix

The preview service has returned 502 on every render since the 2026-08-05 deploy. Confirmed against production at 05:55, 05:56, 06:00, 06:02 and 06:05 UTC - five fresh unused seeds, five 502s, each after ~21s.

Root cause: one line, and it is ours

#154 made the listen port overridable so test/shutdown.test.mjs could bind a free port:

const PORT = Number(process.env.PORT ?? 8080);

But the base image is factoriotools/factorio, and its own image config sets PORT=34197 - Factorio's UDP game port. Read straight out of the registry config blob for the pinned digest:

ENV: PATH=... PORT=34197 RCON_PORT=27015 SAVES=/factorio/saves ...

So in production that expression resolves to 34197, not 8080. The server binds the wrong port, Cloudflare's runtime waits TIMEOUT_TO_GET_PORTS_MS (20s) for something to answer on 8080, then gives up:

Failed to start container: There has been an internal error connecting to the port

Billing corroborates it: placements are born and die within seconds (10 GiB-s in the 05:57 minute, 50 in 06:00).

The fix

Name the override FMW_CONTAINER_PORT. A project-prefixed variable cannot collide with whatever a base image happens to export. 8080 stays the default, matching the Dockerfile's EXPOSE and the Worker's defaultPort in preview-service/worker/src/container.ts.

Why the original check missed it

The local verification was run as env -u PORT node server.mjs - with PORT explicitly unset, which is the single condition under which an inherited-variable bug cannot appear. A test that clears the variable it is defending against proves nothing.

test/port.test.mjs therefore sets PORT=34197 in both cases.

The guard is not vacuous

Established by planting the regression, not by reading it: restoring process.env.PORT makes both new tests fail (pass 0, fail 2).

The second case asserts the 8080 default without depending on 8080 being free on a dev machine - it accepts either listening on 8080 or an EADDRINUSE naming 8080, since both prove the server targeted it.

Testing

Container tests 3/3 pass (render, shutdown, port). Full verify runs in CI here.

After merge

This needs pnpm run preview:deploy to reach production - the deploy is what is currently broken, so merging alone does not restore service.

Then #120's actual measurement can finally run: one 200 render, quiet, wait ~15 min, walk the buckets. Pass = a placementId appearing for a bucket or two then stopping - never "the newest bucket is empty", which is the backfill artifact that manufactured the phantom 8.5-minute tail.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QZvNS2H4cbaZk46ybA7hbj