dev-proxy: rebuild() uses execSync and blocks the supervisor for the length of the build

[debugmcpdev]

Summary

BackendManager.rebuild() in tools/dev-proxy/dev-proxy.mjs runs the build with execSync, so the proxy's event loop is blocked for the whole build — up to DEV_PROXY_BUILD_TIMEOUT_MS (120 s by default, longer on a cold docker rebuild). For that window:

• dev_server_status cannot answer (the request sits in the transport until the build returns);
• the backend's stderr data handlers cannot drain, so a backend that logs heavily during the build fills its pipe (64 KiB on Linux) and blocks on write until the build ends;
• a client disconnect (stdin EOF) is not seen until execSync returns, delaying the shutdown handler's backend.stop() — the orphan class Windows: process tree leaks — dev-proxy and backends outlive their parents; only the proxy tier self-reaps #122 was about.

What would help

rebuild() awaits exec/spawn with the same timeout/maxBuffer and the same sanitized-tail contract for the failure and timeout messages. rebuildThenRestart already runs on the lifecycle queue, so the build stays serialized with restarts; the only change in behaviour is that unrelated requests (status, shutdown) are served while it runs. The two integration cases in tests/integration/dev-proxy-startup.test.ts (failed build, hanging build) pin the message contract.

Raised by the review of PR #745; not done there because it changes the supervisor's concurrency model rather than the lint gate the PR is about.