RFD 1062: A Fly.io toplevel, and the 4090 as a worker node
Problem
RFD 1058 puts the whole stack on one box, with every port bound to 127.0.0.1. That is correct for one operator, and it is not a product. Nobody but the person sitting at this machine can reach it.
RFD 1055 prices three rented GPU options for later, and picks this box’s own RTX 4090 first. RFD 1027 already sized every catalog model at that same 24 GB tier. Renting the tier this box already has is a cost with no matching benefit, for one operator.
Two problems, one shape: this box should keep doing GPU work, and something else needs to be the thing a user reaches.
Decision
This design is abandoned. The split it proposes is not taken.
It rests on this box having an RTX 4090 that no rented tier improves on. The box reports an RTX 3090. Nobody checked the premise, and RFD 1027’s sizing and RFD 1055’s pricing both assumed the same card.
The problem it names does not go away: a stack bound to 127.0.0.1 reaches one operator, which is not a product. Whatever replaces this must answer that, and must name the GPU it measured.
DETAILS.md stays unedited: the full split, the port it reused, and what it left undone.