Add peer-to-peer metrics tracker

[djeebus]

This consolidates metrics into a single struct that does a few things:

• exports metrics to a file called {pid}.json
• watches for other files and reads their metrics.
• when handling incoming requests, check full host metrics

It also creates server.Limiter for checking starting and running limits.

---

Note

Introduces a shared-state manager that aggregates sandbox allocations across processes via PID JSON files and integrates a limiter to cap starting/running sandboxes, wiring both into server and service info paths.

• Shared State (peer-to-peer metrics):
  • Add internal/sharedstate with Manager that writes self allocations to {pid}.json, watches directory via fsnotify, and aggregates Allocations across processes.
  • Integrate with sandboxes map via Subscribe; expose TotalAllocated() and TotalRunningCount().
  • Add tests in internal/sharedstate/tracker_test.go.
• Sandbox start limiting:
  • Add server.Limiter to enforce max running (via featureflags.MaxSandboxesPerNode) and max starting per node; replace semaphore logic in server with sandboxLimiter and error handling.
• Service info metrics:
  • internal/service/service_info.go: switch from iterating local sandboxes to using sharedstate.Manager for CPU/memory/disk and running counts.
• Config:
  • Add SHARED_STATE_DIRECTORY, SHARED_STATE_WRITE_INTERVAL, MAX_STARTING_INSTANCES to cfg.Config.
• Wiring:
  • main.go: instantiate and run sharedstate.Manager; pass to InfoService and server.New; create server.NewLimiter.
  • Minor: rename Google Storage limiter variable for clarity.
• Dependencies:
  • Add github.com/fsnotify/fsnotify to go.mod/go.sum.

^{Written by Cursor Bugbot for commit 835ee09. This will update automatically on new commits. Configure here.}

[linear]

ENG-3153 Export sandbox metrics to a file readable by other orchestrators

[djeebus]

I'm going to pull a piece out to its own PR, and reopen when that's merged.

[sitole]

What about calling it /orchestrator/state so we can put more here in the future?

[sitole]

Ah, this is what I misunderstood during our conversation about host shared metrics. I thought we were asking OS to give us allocated memory, CPU, and disk per whole host and sandbox sums are not returned here. Sorry for the miscommunication here.

[djeebus]

No problem, glad we're on the same page!