We already serve /debug/traces with eztrc — it shows per-request HTTP traces and trcstats already computes p10/p50/p90/p99 percentiles per category. We just don't emit any sync traces into it.

The choke point is DiscoverObjectWithProgress at backend/hmnet/syncing/discovery.go:60. Wrap it with an eztrc trace keyed by the IRI, and mark the stages as they happen:

func (s *Service) DiscoverObjectWithProgress(
    ctx context.Context, entityID blob.IRI, version blob.Version, recursive bool, prog *Progress,
) (blob.Version, error) {
    ctx, tr := eztrc.New(ctx, "sync.discover", string(entityID))
    defer tr.Finish()

    if version != "" {
        eztrc.Tracef(ctx, "version=%s recursive=%v", version, recursive)
    }

    // ... existing body, with Tracef calls at stage boundaries:
    eztrc.Tracef(ctx, "local cache miss, querying peers")
    // ...
    eztrc.Tracef(ctx, "local peers: %d found, %d synced ok", prog.PeersFound.Load(), prog.PeersSyncedOK.Load())
    // ...
    eztrc.Tracef(ctx, "DHT lookup start")
    // ... etc.
}

Per-peer sync gets its own nested trace inside syncWithPeer at backend/hmnet/syncing/syncing.go:351:

func (s *Service) syncWithPeer(ctx context.Context, pid peer.ID, eids map[string]bool, ...) error {
    ctx, tr := eztrc.New(ctx, "sync.peer", pid.String())
    defer tr.Finish()
    // ... existing body, with Tracef at RBSR rounds and blob download batches.
}

Once this is in, /debug/traces shows sync.discover and sync.peer as new categories. trcstats gives us p50/p90 for free. No new endpoints, no new ring buffer, nothing to garbage-collect.

The only judgment call is how many stage marks to emit. Too few and we can't tell where time went; too many and the trace view gets noisy. Starting points: entry, local-cache-hit shortcut, peer list size, each peer sync start/end, DHT lookup start, DHT peer sync start, exit. Tune once we see real traces.

net/http/pprof is already mounted on the daemon's debug listener (backend/cmd/seed-daemon/main.go:15, registered in backend/daemon/http.go:147). That means GET http://localhost:56001/debug/pprof/profile?seconds=30 already works today — it just has no UI.

Add a small block in the desktop app's Advanced settings (frontend/apps/desktop/src/pages/settings.tsx, inside DeveloperSettings) with three buttons:

• Capture 30s CPU profile → /debug/pprof/profile?seconds=30

• Heap snapshot → /debug/pprof/heap

• Goroutine dump → /debug/pprof/goroutine

The click handler reuses the exact pattern in frontend/apps/desktop/src/save-cid-as-file.tsx: a tRPC mutation in app-api.ts that uses Electron's net.request to stream the response, dialog.showSaveDialog to pick a target, and fs.writeFileSync to persist. Default filename like seed-cpu-<iso-timestamp>.pprof.

Optional finisher: after save, shell.openPath the containing folder so the user can drag the file into an issue or open it with go tool pprof -http.

That's the whole feature on the frontend. Daemon gets zero new code.