Advise: add-power-thermal

ADVISE: accept READER: sol-arch-review SPAWN: /Users/dukejones/work/ClientProjects/AllSystemsGo/AICamera/.spawns/add-power-thermal-1789036563-71978-5d435db3

Independent take (blind pass)

Written from the Why, living hardware-kit spec, docs/POWER.md, and its T4000/thermal references before opening design.md or tasks.md.

1. Pin record continuity as the protected load: any power or thermal shortfall sheds AI before capture, encode, metadata, or storage paths.
2. Pin the production baseline to the T4000 SOM at 70 W default, and keep the T5000 AGX kit’s 40–130 W lab behavior explicitly separate.
3. Treat 90 W TMP and 75 °C TTP as throttle and absolute interface limits, not nominal design targets; require operating headroom.
4. Refuse any claim that the resin first article validates production cooling, sealed operation, fanless 70 W, or sun-load performance.
5. Require separate, unambiguous power trees for Thor and the PoE switch, with no PoE backfeed path into the Thor carrier.
6. Budget two 802.3at satellites at about 60 W at the PSE while also accounting for the switch’s own 15–30 W supply load.
7. Require the design to state where power and temperature are sensed and what deterministic control action enforces record-first degradation.
8. Require staged verification: rail isolation and load shedding first, then TTP/ambient measurements on the real T4000 carrier and cooler stack.
9. The central tradeoff is full 70 W AI performance versus quiet compact operation; record-only power must remain viable when cooling cannot sustain AI.
10. Keep the AGX-kit enclosure useful for fit and workflow bring-up, but prevent its dimensions, fan, or brick from becoming production evidence.

Comparison

The change answers the hardware-buying decisions. The split-power-tree SHALL separates Thor from satellite PoE and forbids backfeed. Its sag scenario makes AI the shed load while recording continues. The thermal SHALL carries the T4000’s 70 W default, 90 W maximum TMP, and 75 °C TTP ceiling; the enclosure scenario rejects resin as the thermal path. design.md also keeps the AGX-kit first article distinct and budgets the PoE switch separately.

The delta does not name sensor placement, trip hysteresis, or a qualified operating margin below 75 °C. Those are real validation concerns, but they do not leave the architecture indeterminate: the observable degradation order and absolute limits are normative, and measured T4000 thermal qualification is explicitly outside this change. The implementation must not reinterpret 90 W as a nominal sustained target or use the resin fit-check as thermal evidence.

Steelman for send-back

“Thermal or power headroom is gone” is not itself a measurable trigger, and “designed for 70 W default and 90 W TMP” could be misread as permission to live at the throttle ceiling. A more exhaustive delta would specify sensing, margins, and the proof load.

That does not justify rejecting this architecture pass. Exact thresholds depend on the carrier, PSU, cooler, ambient, and measurements not yet available. Adding invented values now would create false precision. The current requirements preserve the decisions that must survive folding: separate supplies, no backfeed, record-first shedding, the correct T4000 limits, and no resin thermal claim.

Notes

• The eventual verification should demonstrate that AI shedding does not interrupt picture, timecode, lens-sidecar, or storage paths.
• Qualify TTP margin with the real T4000 + Rogue-T5 + cooler stack before sealed-body or sun-load claims; the 75 °C figure is a maximum, not the desired control point.
• The AGX kit remains a T5000 lab brick with its own fan and power behavior. Its resin enclosure proves fit and ventilation only.

Verdict rationale

Accept. The delta is lean but internally consistent with the cited NVIDIA limits and the project’s record-first invariant. It prevents the two expensive architectural mistakes this node exists to prevent—combining Thor and PoE power, and treating resin or the AGX lab kit as proof of a production T4000 thermal solution—without pretending unmeasured thermal details are settled.