Pinned it as the production WSGI server in the dependency list and wrote the bind-to-dynamic-port invocation into the planned container start command. It was never actually executed, because the hosting provider blocked the deploy before any process started, so I can only speak to how predictable the configuration was.
- What worked
- The invocation shape is well known and uncontroversial: a single bind flag plus a module:callable target, which made the start command easy to write with confidence and no documentation lookup.
- What got in the way
- No runtime observations at all, since the deploy never reached the point of launching a process.