feat: NV_ERR_NO_MEMORY on GPU panel (slow journal scan)

[MiaAI-Lab]

Why this replaces #40

#40 put an NV_ERR_NO_MEMORY counter on CpuPanel and ran journalctl -k | grep -c on the unified-memory poll (~2s). CpuPanel is gone, and a full kernel-journal scan every 2s is not a metric.

What this does

• Count NVRM NV_ERR_NO_MEMORY lines in the kernel journal since boot
• Attach it to GPU metrics and show it on the GPU panel when the count is > 0
• Scan at most once per POLL_INTERVAL_NVERR (default 60s), cached — not on the 2s GPU/memory loop
• Local Docker uses host nsenter (same as nvidia-smi); remotes use SSH
• No per-browser localStorage reset; the number is journal-since-boot
• Parser unit tests only (no --experimental-test-module-mocks)

Test plan

• node --test server/collectors/__tests__/SystemCollector.nvErr.test.js
• npm run typecheck
• On a Spark/host with no NVRM errors: GPU panel unchanged (row hidden)
• After NV_ERR_NO_MEMORY in journalctl -k: row appears on GPU panel, not a CPU panel

[vincenzopalazzo]

ACK — NV_ERR_NO_MEMORY on GPU panel

Production Readiness: GO

Minimal. Cache + POLL_INTERVAL_NVERR is the load-bearing choice vs #42. Take this version. Merge.

— Vincent

[vincenzopalazzo]

Already landed via #78 (Batch A ACK merge onto main). Closing this one as superseded — thanks for the contribution!

— Vincent

[vincenzopalazzo]

Correction: #78 is ready but not yet merged (needs MiaAI-Lab write access). Once #78 is rebase-merged, please close this PR as superseded. Thanks again!

— Vincent

[vincenzopalazzo]

Already merged via #78. Thanks for the contribution!

— Vincent