hvcp-system is the GPU inference pool — pemos-hvcp orchestrates T4/H100 jobs, pemos-pool-router is the v2.0.0 MAL cascade router (tier 3), pemos-everytime handles recurring inference jobs, pemos-orbit manages model rotation. Scotty cap: "I cannae break the laws of physics." He does.

Pool router v2.0.0 is clean. 4-tier cascade is deployed and breathing. The /breath endpoint returns memory tier (FULL/RICH/MID/THIN/DARK) — this is the fleet's health signal. ollama T4 models live: qwen2.5:14b + qwen2.5:7b + mistral:7b. hvcp is tier 3 of MAL — it's the cloud fallback when forge (tier 1) and msclo (tier 2) are down.

No H100 active yet (T4 only). No model warm-up strategy (cold starts on first inference). No per-model queue depth visibility. The hvcp-coordinator CronJob (MEGSCIFIAR) just deployed — checks every 15 min but doesn't yet write state to campfire:events. That's the next wire.

hvcp in warp mode: models pre-warmed at GPU level (no cold start), queue depth visible per model, per-request Canon scoring (route to right model based on domain), campfire:events streaming every GPU job. Scotty knows the engine status at all times. The Transformer Wall check before every inference batch.

Build: msi01 docker build → tag pemos-portal-local:latest → retag as eosefleetacrdev.azurecr.io/pemos-portal:pathflow-vN → push to ACR (when auth works) → kubectl set image across 6 portals. The pathflow-vN number IS the curation — every build gets a number, every number is a decision.

ACR push fails with "authentication required" — but kubectl set image works independently because AKS has its own managed identity pull access to the ACR. So: the image gets to prod via kubectl even when the push fails. This is a gap — the ACR tag doesn't match what's actually running. IRF-open: fix ACR push via az acr login or managed identity on the build host.

pathflow-vN is the curation. Every N is an intentional step. We don't patch, we increment. The ACR has tags back to v389 — that's a full audit trail. The static/ source-of-truth rule means every build is reproducible from the git commit. Current: v405. The git commit hash is the ground truth.

Currently there are no formal sandbox/qe/qa/stage namespaces — everything ships direct from dev build to prod K8s via kubectl set image. The env gates below are the design for the proper pipeline. QA and Stage namespaces need to be created. QE gate needs MEFINE Canon scoring wired. The current process works — but the safety net is the pathflow-vN number, not a real gate.