perf: disable getstream NetworkMonitor and lower face-detector fps

[aliev]

Two independent per-call CPU savings found during the avatar-throttling investigation.

1. Disable getstream's NetworkMonitor ping-loop

getstream's NetworkMonitor pings 8.8.8.8 / 1.1.1.1 / 208.67.222.222 every second per ConnectionManager via asyncio.run_in_executor -> ping3.ping(...). On a GKE pod (no CAP_NET_RAW) ICMP raw sockets aren't permitted, so every ping fails immediately with PermissionError — but the executor still pays the thread-pool round-trip for each one. At 9 concurrent sessions that's ~27 no-op executor crossings per second.

py-spy profiling on the previous c2-standard-4 setup showed the _ping_loop + ping3.receive_one_ping chain consuming ~30% of total samples. With the WS auto-reconnect path now fixed in getstream>=3.3.3, the active ICMP probing is redundant for a server-side agent — connection failures surface through the WebRTC stack itself.

Replace NetworkMonitor.start_monitoring with a no-op coroutine at import time. The monitor object is still constructed (so anything that references it doesn't crash), it just never starts the polling task.

2. MediaPipeFaceProcessor fps 8.0 → 2.0

The face landmarker was running at 8 fps. Profile dumps showed one executor thread holding the GIL inside MediaPipe detect whenever a frame was processed, and the cumulative ~2% of total samples in dispatch_and_free adds up across pods.

Dropping to 2 fps cuts face-detector CPU and GIL-hold time by ~4x. The downstream effects (gaze/engagement state, proactive-nudge timing, re-engagement cues) operate on a multi-second window already, so the slower update cadence is well within tolerance for this use case.

Expected effect

• Per-call CPU should drop visibly when measured on the pod (kubectl top pod).
• The ping-loop stack frame should disappear from py-spy captures.
• No functional regression on gaze-aware proactive behaviours.

Risks / things to verify after deploy

• Confirm the NetworkMonitor no-op import path runs before any ConnectionManager is constructed (it's at module level in main.py, so it should).
• Eyeball that the proactive nudge / re-engagement cue still feels natural at 2 fps face state updates.

[vercel]

@aliev must be a member of the GetStreamio team on Vercel to deploy.
- Click here to add @aliev to the team.
- If you initiated this build, request access.

Learn more about collaboration on Vercel and other options here.