Per-service restart and lifecycle management

[raamsri]

Problem

vctl service up/down manages all services as a single unit via lifecycle.start/stop. There's no way to restart an individual service (e.g., restart a crashed pipecat worker without restarting the inference server).

When a service crashes (e.g., pipecat worker dies from a GStreamer bug), the only recovery is to restart everything or manually SSH in and run the start command.

Proposed

Add per-service lifecycle commands:

vctl service restart <service-name>
vctl service logs <service-name>

This requires velocity.yml services to declare their own start/stop commands:

services:
  atoms-worker:
    path: ./atoms
    start: $VENV/python main.py --worker-mode warm
    stop: pkill -f "main.py --worker-mode"
  inference:
    path: ./inference
    port: 8100
    start: $VENV/uvicorn app.main:app --host 0.0.0.0 --port 8100 --workers 2

Currently services only declares path, port, and routes — the actual process management is embedded in the monolithic lifecycle.start shell script.

Context

• Pipecat worker crashed during audio mixing (GStreamer stack smash), leaving all subsequent calls with no available worker
• Recovery required manually SSH-ing and running the exact start command with correct paths
• atoms-platform uses turbo which manages child processes, but pipecat uses nohup background processes with no supervision