Network sidecar crashes on reconnect: SidecarConfig required for NetworkNodeAdapter.start()

[Hafuyya]

Summary

When HYPERSPACE_NETWORK_SIDECAR=1 (default), any network disconnect triggers a reconnect path that crashes with:

Fatal error: Error: SidecarConfig required for NetworkNodeAdapter.start()

This forces a full process restart every time the sidecar loses connectivity. The only workaround is HYPERSPACE_NETWORK_SIDECAR=0 (in-process networking), which introduces a separate problem: training gradient uploads block the libp2p event loop, causing heartbeat timeouts and ~10-minute crash cycles.

Root Cause

The CLI has two this.node.start() calls:

1. Initial start (in try/catch) — passes sidecarConfig ✅
2. Reconnect path — calls await this.node.start() without sidecarConfig ❌

NetworkNodeAdapter.start() requires sidecarConfig to spawn hyperspace-network, so the second call always throws.

Reproduction

1. Start with HYPERSPACE_NETWORK_SIDECAR=1 (default)
2. Wait for any network disconnect (peer timeout, transient connectivity loss, etc.)
3. Observe crash: SidecarConfig required for NetworkNodeAdapter.start()
4. Service restarts, works until next disconnect, crashes again

Tested on CLI versions v5.37.3 through v5.39.6 — all affected.

Impact

• With sidecar (default): Node crash-loops on any disconnect event
• Without sidecar (workaround): Training gradient upload + FedAvg aggregation blocks the single-threaded event loop, causing GossipSub heartbeat timeouts → peer disconnections → node restart every ~10 minutes

This means there is currently no stable configuration for miner nodes that participate in training. The sidecar fix would resolve both issues since networking would run in a separate process, unaffected by training I/O.

Environment

• OS: Ubuntu 24.04, Linux 6.17.0
• GPU: RTX 5090 32GB
• CLI: v5.39.2 (also tested v5.37.3, v5.37.4, v5.37.6, v5.39.6)
• Chain: v1.7.4
• Role: miner

Suggested Fix

Pass sidecarConfig to the reconnect start() call, or cache it from the initial start so NetworkNodeAdapter can reuse it on reconnect.