Jetson AGX Orin 64GB — board state across the 2026-08-25 Lightning spikes
=========================================================================

PRE-SPIKE INVENTORY (captured from `docker ps`, before the first stop):

  model-gear-gateway        Up 2 days (healthy)     ~42.67 MiB
  llamacpp-cortex           Up 2 days               ~27.63 GiB
  prod-worker-1             Up 2 days               ~20.45 MiB
  model-gear-vllm-rerank    Up 2 days (healthy)     ~5.355 GiB
  model-gear-vllm-embed     Up 2 days (healthy)     ~5.041 GiB

  Instrument: `docker stats --no-stream`. NOTE (frame claim c32/h13): on this
  Tegra board `nvidia-smi` reports memory.used [N/A] — there is no GPU-memory
  query. Host-side readings are unified-memory figures; the profile's own
  ~33 GiB figure for llamacpp-cortex disagrees with the 27.63 GiB above.
  Prefer an engine's own boot-log accounting over any host-side reading.

STOPS PERFORMED, and why:

  llamacpp-cortex        stopped — a 30B Q4 needs ~18 GiB; only ~8 GiB was free.
                         Required for EITHER lane, not just vLLM.
  model-gear-vllm-embed  stopped — headroom for util 0.7 (~42.9 GiB).
  model-gear-vllm-rerank stopped — same.
  model-gear-gateway     stopped — the vLLM recipe uses --network host and
                         binds :8000, which the gateway held.
  prod-worker-1          LEFT RUNNING throughout (unrelated workload, ~20 MiB).

POST-SPIKE STATE — DELIBERATELY NOT RESTORED (operator decision, 2026-08-25):

  The four lanes above remain STOPPED by explicit operator instruction. The
  board is being repurposed: `associate` (Nemotron 3.5 Lightning, vLLM lane,
  GO per docs/evidence/2026-08-25-spike-lightning-vllm-orin.txt) becomes this
  box's generate lobe, so restoring the llama.cpp GGUF cortex — measured at
  2.61 tok/s against associate's ~78-81 tok/s — would be restoring the lane
  this work replaces.

  This is an ABANDONMENT, not an oversight. It satisfies the second acceptance
  criterion of plan task t1 ("...OR the transcript explicitly declares them
  abandoned and why"), and it resolves plan risk r1: with no restore, task t8's
  budget measurement needs no second outage.

  Consequence to state plainly: until the associate lane ships, this box serves
  NO local generate role, NO embedder and NO reranker, and its gateway is down.
  Mesh peers addressing this box will fail rather than be served. Restore at any
  time with:

    docker start llamacpp-cortex model-gear-vllm-embed model-gear-vllm-rerank model-gear-gateway

SPIKE CONTAINERS: both removed. `nemotron35-vllm` (vLLM) and
`lightning-llamacpp-spike` (llama.cpp) no longer exist; port 8000 is closed and
the unauthenticated tailnet exposure recorded in c30/c46 has ended.
