Description
AI Factory racks draw tens to over a hundred kilowatts. Air-cooled PCIe GPU racks are metered at the PDU over SNMP. Liquid-cooled NVL72 racks report rack power from their power shelves over the out-of-band network and depend on a coolant distribution unit, read over SNMP or Modbus, for secondary-loop supply and return temperature, flow, pressure, pump state and leak detection. No CDU vendor ships a Prometheus exporter, so collection is snmp_exporter or a Modbus exporter, and a cooling fault shows up first as GPU thermal throttling. Facility-side BMS telemetry is a separate follow-up.
Acceptance Criteria
- #1 PDU metering is emitted through the SNMP exporter construct using real PDU modules
- #2 CDU telemetry is emitted through a real SNMP module or a real Modbus exporter with provenance; register maps or objects that cannot be sourced are PENDING in cantfind.md
- #3 The power shelf collection path is sourced from documentation or recorded as PENDING
- #4 Physics are consistent: rack power tracks the sum of node power plus overhead, and CDU heat load tracks rack power
- #5 Failure modes cover a CDU pump failure that raises coolant temperature and causes DCGM thermal throttling on that rack’s GPUs, a leak detection, and a PDU branch approaching its limit
- #6 Every metric, label, attribute and log field is sourced from vendor documentation or a capture at a pinned version and recorded with provenance in the named signals/ area file; anything unconfirmed is a PENDING in cantfind.md, never an invented name
- #7 Rates, counts and cardinality in any shipped blueprint are plausible for a real deployment, with the real-world basis stated in a comment; nothing is raised to force emission
Definition of Done
- #1 just check (fmt-check, lint, gen-check, env-check, docs-check, test, race, hygiene, ui-check, compose-check, helm-test, lab-check, signal-fidelity)
- #2 just gen (only if a blueprint field, construct/workload config struct, or a skill under plugins/synthkit/skills/ changed)
- #3 just dump — inventory diffed against signals/
Implementation Notes
loop47: admitted R14 facility vendor-contract research; implementation waits for fixture/DCGM/SNMP/Redfish and parked BCM source contract
loop47: final facility research pins snmp_exporter v0.30.1 infrapower_pdu and APC bank current/state; no authoritative CDU model MIB/generated profile or Modbus register map and no power-shelf telemetry protocol/schema. NVIDIA guide documents ipminet reachability only, not IPMI protocol. Resume with those exact vendor artifacts and dependencies. Implementation 0; review-repairs 0; infra retries 0; not rejected. R15 remains untouched To Do because its .14-landed admission boundary is unmet.