oj v0.2.19

Fixed

• Editing a file that a plugin's resolveId answers for (a framework's generated file, an alias target) now hot-updates. Such files were compiled by the plugin route itself, never watched and absent from the module graph, so saving them changed nothing until a manual reload; a resolved id naming a plain app file now goes through the normal module pipeline, which still asks the plugin's load before reading the disk and runs the transform hooks, as Vite does for every real file. Module-graph nodes in the plugin host also carry the module's served URL (root-relative, or /@fs outside the root) instead of its filesystem path, so HMR updates built from them name URLs the browser actually imported, and urlToModuleMap/getModuleByUrl answer by that URL as in Vite.
• Plugin-spawned children (a Cloudflare plugin's workerd, a framework's sidecar) no longer outlive the dev server. The shutdown sweep now covers every path that used to strand them: the start and ssr serve modes install the same shutdown handler as plain dev (they exited on a bare signal before); the embedded engine's own signal handling, which terminates the process before any parked shutdown task can run, sweeps children on its exit path too; and the handler is installed before boot begins, as Vite installs its SIGTERM listener inside createServer before any configureServer hook can spawn, so a signal that lands mid-boot still sweeps. Against SIGKILL, where no handler runs at all (an OOM kill, a supervisor escalating after a grace period), the child registry is persisted per server process with pid plus kernel start-time identities; the next boot reaps the previous server's still-identified children before any engine spawns, the way Bazel- and Gradle-style daemon registries recover, and never signals a recycled pid.