fix(manager): close idle inbound connections

[alongubkin]

Why

axum::serve sets no timer, so it never closes an idle keep-alive connection. Any client that opens connections and forgets them pins hyper's per-connection read buffers in the manager forever. In production that was ~8,400 idle connections and ~1.6Gi per replica after three hours, from a client bug (the demo's per-tick OTLP exporters, fixed separately). The manager should not depend on every client being well behaved.

Linear: ALIEN-1152

What

server.rs serves connections through hyper-util's auto::Builder with a tokio timer and header_read_timeout. hyper applies that timeout between requests on a keep-alive connection too, so a connection that sends no request for 90s is closed. That's above the ALB's 60s idle timeout, which AWS requires so the balancer never reuses a connection the target has just closed. Upgrades (debug WebSockets) are still served via serve_connection_with_upgrades, and accept errors (fd exhaustion) are logged and retried as axum::serve did.

Two new crate dependencies, both already in the workspace: hyper and hyper-util.

Validation

• New test idle_keep_alive_connection_is_closed_after_the_timeout: raw TCP to a server with a 300ms idle timeout. Two requests inside the window are served on the same connection (keep-alive still works); then the client goes silent and the server closes the socket (read returns EOF). It fails on main's behavior.
• cargo test -p alien-manager --lib: 291 passed.
• cargo fmt --check clean. Clippy via CI.

After deploy, ls /proc/1/fd | wc -l on a manager pod should stay in the low hundreds even with a leaking client.

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 5/5

[High risk] Replaces the HTTP server implementation with custom connection handling.

The PR appears safe to merge based on the reviewed changes.

Summary

The PR replaces the manager’s server loop with an HTTP/1 server configured to close idle inbound connections, and adds coverage for silent and idle keep-alive sockets.

_{Reviews (2) · Last reviewed commit: "fix(manager): serve HTTP/1 only so the i..."}

[alongubkin]

@greptileai re-review: switched to HTTP/1-only serving so silent and idle connections are both closed; test reads full responses and covers the silent case.