BASELINE — cortex from the Jetson AGX Orin pins to ONE peer
============================================================
Date:      2026-08-30
Box:       Jetson AGX Orin 64 GB (sm_87), shape: associate + hand + embedder +
           reranker; cortex NOT hosted (PRIMARY_FEASIBLE=false)
Peers:     Spark  http://spark.tail0be7e0.ts.net:8001
           Thor   http://thor.tail0be7e0.ts.net:8000
Purpose:   the PRE-change baseline for peer-only replica pools
           (docs/specs/2026-08-30-peer-only-replica-pools.md, claim c13/h13).
           Captured BEFORE any code change landed, per the #108 rule.

This is a BASELINE, not an acceptance run. It records what the deployed
gateway does today so the later acceptance transcript has something honest to
be compared against. Nothing here was restarted, rebuilt, or reconfigured for
the measurement; the only change on the box that day was
PRIMARY_PEER_PROXY=false -> true (see "Deploy record" below), made before
these numbers were taken.


(0) DEPLOY RECORD
-----------------
~/.lobes/.env on the Orin, cortex lines (secrets elided):

    PRIMARY_FEASIBLE=false
    PRIMARY_PEER_ORIGIN=http://spark.tail0be7e0.ts.net:8001
    PRIMARY_PEER_PROXY=true          # flipped from false on 2026-08-30
    PRIMARY_PEER_API_KEY=<spark's inbound GATEWAY_API_KEY>
    # no PRIMARY_PEER_ORIGINS / PRIMARY_PEER_API_KEYS declared

Gateway container recreated once to pick that up
(`docker compose up -d --no-deps gateway`); model-gear-vllm-associate was not
touched (Up 4 days, healthy, answering 200 throughout).

Before the flip, `model=cortex` on this box answered:

    HTTP 404
    {"error": {"message": "The model `cortex` is not feasible on this machine
     ... It is hosted by the peer at `http://spark.tail0be7e0.ts.net:8001` —
     address that box directly; this gateway never proxies requests to peers.",
     "type": "role_infeasible", "code": "role_infeasible", ...}}


(1) BOTH PEERS HOST CORTEX, AND ALREADY POOL WITH EACH OTHER
-------------------------------------------------------------
GET /capabilities on each peer, cortex entry:

    spark: feasible=True ready=True context=262144  proxied=None hosted_by=None
      replicas:
        http://vllm-primary:8000                local ready=True compat=True
        http://thor.tail0be7e0.ts.net:8000      peer  ready=True compat=True
    thor:  feasible=True ready=True context=262144  proxied=None hosted_by=None
      replicas:
        http://vllm-primary:8000                local ready=True compat=True
        http://spark.tail0be7e0.ts.net:8001     peer  ready=True compat=True

Both serve unsloth/Qwen3.8-27B-NVFP4 at max_model_len 262144, runtime vllm,
quantization compressed-tensors. The #199 pool between them is armed and
mutual. The Orin is the box that cannot reach it.


(2) THE DEFECT — 4 CONCURRENT REQUESTS, ALL PINNED TO ONE PEER
---------------------------------------------------------------
4 concurrent POST /v1/chat/completions with model=cortex, max_tokens=160,
temperature=0, issued to the Orin's own gateway:

    req1 served_by=http://spark.tail0be7e0.ts.net:8001 route_reason=none
         completion_tokens=155 elapsed=3.594s
    req2 served_by=http://spark.tail0be7e0.ts.net:8001 route_reason=none
         completion_tokens=160 elapsed=3.747s
    req3 served_by=http://spark.tail0be7e0.ts.net:8001 route_reason=none
         completion_tokens=154 elapsed=7.032s
    req4 served_by=http://spark.tail0be7e0.ts.net:8001 route_reason=none
         completion_tokens=155 elapsed=7.032s

    wall_clock_total = 7.080 s
    completion tokens = 624
    AGGREGATE = 624 / 7.080 = 88.1 tok/s

    per-peer split: Spark 4/4 (100%), Thor 0/4 (0%)
    X-Lobes-Route-Reason: absent on every response (no placement happened)

The 3.6 s / 7.0 s split is the queue: two requests complete in the first
batch, two wait for the second. The Spark's PRIMARY_MAX_NUM_SEQS is 2.

Peer load sampled immediately after the burst:

    spark /status primary: health=ok running=0 waiting=0
                           requests_succeeded=361
    thor  /status primary: health=ok running=1 waiting=0
                           requests_succeeded=573

The Thor was healthy and serving throughout. It received none of this box's
four requests.


(3) THE ADVERT IS ALSO WRONG TODAY
-----------------------------------
GET /capabilities on the Orin, cortex entry:

    {'feasible': False, 'ready': False, 'context': 1048576,
     'proxied': True, 'hosted_by': 'http://spark.tail0be7e0.ts.net:8001',
     'loaded': True}

Both fields disagree with the peers that actually serve it:

    ready:   False here, True on both peers
    context: 1048576 here, 262144 on both peers

1048576 is this box's own PRIMARY_MAX_MODEL_LEN mirror, not a window any
replica serves — the same class of defect #220 fixed for the single-peer
relay, reappearing because a box with several declared replicas has no single
peer whose advert can be relayed.


(4) WHAT THIS BASELINE COMMITS THE ACCEPTANCE RUN TO
-----------------------------------------------------
Success signal c23 measures against exactly these numbers:

  * per-peer split must show BOTH origins (neither 0%, neither 100%);
    baseline: 100% / 0%
  * aggregate must exceed 88.1 tok/s by at least 40% (>= 123.3 tok/s)
  * with one peer's gateway stopped, 100% of requests must still answer 200
  * ready/context must match what the replicas actually serve

Per honesty condition h22, any target not met is reported as NOT met in the
acceptance transcript; none is softened after the run.
