caveman-shrink never forwards SIGTERM/SIGINT to its spawned upstream server, orphaning it #742

What happen

caveman-shrink (the MCP proxy that wraps an upstream MCP server and compresses tool
descriptions) never forwards SIGTERM/SIGINT to the upstream process it spawned. If
something tears down the caveman-shrink process directly with a signal (an MCP host
ending a session, disabling a tool, reloading config) instead of killing a whole
process group, the wrapped upstream server is left running, reparented to PID 1,
indefinitely.

src/mcp-servers/caveman-shrink/index.js only registers handlers for the child's own
exit/error events:

const upstream = spawn(args[0], args.slice(1), getSpawnOptions());

upstream.on('error', err => { ... });
upstream.on('exit', (code, signal) => { ... });

There is no process.on('SIGTERM', ...) / process.on('SIGINT', ...) on the parent to
forward the signal (or kill the child) before exiting.

Expected

Sending SIGTERM to the caveman-shrink process should also stop the upstream child it
spawned, not leave it running orphaned.

Repro

# fake long-running upstream (never exits on its own)
cat > /tmp/fake-upstream.js <<'EOF'
process.stdin.resume();
setInterval(() => {}, 1000000);
EOF

node src/mcp-servers/caveman-shrink/index.js node /tmp/fake-upstream.js &
WRAPPER_PID=$!
sleep 1
kill -TERM $WRAPPER_PID
sleep 1
ps -o pid,ppid,stat,args

Verified today in a clean node:20-alpine container against current main
(0d95a81): the wrapper process becomes a zombie, and
the fake-upstream process is still alive with PPID 1 (reparented, orphaned), never
having received a signal.

Platform

• Claude Code
• Codex
• Other: any MCP host wrapping a server via caveman-shrink (not agent-specific)

Version / install method

main @ 0d95a81d, ran the script directly from the repo (not through the packaged
installer).

Happy to send a PR (forward SIGTERM/SIGINT to the child, mirroring the existing
upstream.on('exit', ...) handling) if useful.