fix(runtime): clamp fixed-port applications to one worker when reusePort is not available - #5076

Fixes #5070

Stacked on #5075 (restore portAssignment: perWorkerIncrement) — review only the last commit; retarget to v4 once #5075 lands.

Summary

The guard that clamped a multi-worker application to a single worker on platforms without SO_REUSEPORT (macOS, Windows) was removed with the entrypoint in #5014, because it was gated on application.entrypoint. On those platforms an application with a fixed server.port and workers > 1 failed to start its additional workers with a raw (and, for @platformatic/node factories, delayed) error instead of degrading with a warning.

Changes

• Restore the clamp, gated on the listener configuration (packages/runtime/lib/runtime.js, #setupApplication): the listener is owned by the capability, so when reusePort is not available and the application has workers > 1 (or dynamic scaling), the runtime sets up the first worker, reads its configuration via ITC (getApplicationConfig) and, if server.port is a fixed port and server.portAssignment is not perWorkerIncrement, logs a warning naming the OS limitation and the portAssignment alternative, sets workers to { dynamic: false, static: 1 } (which also disables the dynamic scaler for that application) and skips the other workers. Applications with perWorkerIncrement, port 0, or no managed listener are unaffected.
• Meaningful error when workers of the same application collide anyway (e.g. after a manual updateApplicationsResources scale up, or with reuseTcpPorts: false):
  • packages/runtime/lib/worker/itc.js: the start handler now sends the failure as a plain object, so error properties such as port/address survive the structured clone (previously the error.port check in #startWorker could never match through ITC).
  • #startWorker reports the new PLT_RUNTIME_WORKER_EADDR_IN_USE (WorkerAddressInUseError) when the port is held by another worker of the same application, keeps PLT_RUNTIME_EADDR_IN_USE for another application, and leaves the original EADDRINUSE when the owner is an external process (instead of the odd "another process" and "app" message). Both codes are non-retryable.
  • (The related @platformatic/node unhandled-rejection fix, which turned a listen error into a 30s start timeout, moved to fix(runtime): restore per-worker port assignment (portAssignment: perWorkerIncrement) #5075 since its Windows CI needed it.)
• Docs: workers/reuseTcpPorts sections and errors.md.

Tests

• packages/runtime/test/multiple-workers/no-reuse-port.test.js: simulates the missing reusePort feature (the runtime runs in-process, so features.node.reusePort is patched) and covers the clamp with static workers, the dynamic scaling case, no clamp with perWorkerIncrement and with an ephemeral port, plus the same-application collision error (forced with reuseTcpPorts: false, so it runs on Linux too).
• packages/runtime/test/ports.test.js now asserts the exact PLT_RUNTIME_EADDR_IN_USE error for two applications on the same port.

Assisted-by: Claude Code:claude-fable-5