Fix connection slot leak when Chrome crashes while client is idle

[marcosvrs]

Problem

When a Chrome process crashes (OOM, segfault, etc.) while the connected client is idle, the connection slot is permanently leaked and never reclaimed. Over time this exhausts MAX_CONCURRENT_CHROME_PROCESSES and all new connections are rejected.

Root cause

The proxy creates two bridging tasks:

taskA = asyncio.create_task(hereToChromeCDP(...))   # Chrome -> Client
taskB = asyncio.create_task(puppeteerToHere(...))    # Client -> Chrome
await taskA
await taskB

When Chrome crashes:

1. taskA detects the broken CDP WebSocket (ConnectionClosed) -> completes
2. taskB blocks on async for message in chrome_websocket -- the client WebSocket is still alive, so it waits for a message that never comes -> hangs forever
3. await taskB never returns -> handler never exits -> websocket.wait_closed() callback never fires -> stats_disconnect() never runs -> semaphore never released

The slot is only freed if the client happens to send another message (triggering a send to the dead Chrome ws) or disconnects. If the client is idle -- common for long-running browser sessions -- the slot is gone permanently.

Evidence

After 3 days of operation with MAX_CONCURRENT_CHROME_PROCESSES=3:

• Stats: Active count 3 of max 3 with only 2 live Chrome processes
• Network: 4 CLOSE_WAIT connections to dead Chrome debug ports (11845, 11680, 11854, 11875) that no longer exist
• Processes: Zombie Chrome entries in the process table from killed-but-not-reaped processes

Fix

1. Cross-task cancellation (primary fix)

Replace sequential await taskA; await taskB with asyncio.wait(FIRST_COMPLETED):

done, pending = await asyncio.wait(
    [taskA, taskB],
    return_when=asyncio.FIRST_COMPLETED
)
for task in pending:
    task.cancel()

When either side of the proxy dies, the other is cancelled immediately. The handler returns, the WebSocket closes, and stats_disconnect() fires normally to release the slot.

2. Process reaping (secondary fix)

Add chrome_process.wait() after SIGKILL in cleanup_chrome_by_pid() to reap dead Chrome processes instead of leaving zombies in the process table. Related: #46

Testing

Deployed the patched server.py into a running dgtlmoon/sockpuppetbrowser container. After 9+ hours and 108 processed connections:

• Active count stays at 0 when idle (previously accumulated leaked slots)
• All Chrome processes properly reaped (reaped successfully in logs)
• Zero CLOSE_WAIT connections to dead ports
• Health checks passing continuously