conformance: measure the runtime through an ssh tunnel, not the Lium port proxy

[anderdc]

Three of the four failures since the last conformance run were the harness, not the runtime — and the one pass it produced (#1740) carried a value that would have failed every honest miner.

R6 failed twice on 91.224.44.226 with an image that passed on 91.224.44.81 on 8/31. Rented that host with ssh and measured: a 2 s HTTP handler answers 16 parallel requests in 2.4 s … 38.5 s through the Lium-mapped port and 2.2 s … 2.2 s on-box — the port proxy caps concurrent connections at ~5 per pod (5 clean, 6+ serialise). That serialises the 16-wide overload and speed bursts: a false R6 failure, and an aggregate_decode_tps of 25.9/46.9 against 263 — the number the validator has priced tokens from since #1737 ($0.69/M would have read $7.5/M). Through one ssh connection the same 32 parallel finish in 2.5 … 3.6 s, so the checker now reaches the runtime over ssh -L 18080:127.0.0.1:8080. Lium's SDK registers only ~/.ssh/id_ed25519 with a pod (it ignores LIUM_SSH_KEY_PATH and the config file), so the workflow generates the key there. Run 33532638861 on the same host: CONFORMANT, R6 16×200 in 32.4 … 32.9 s.

attest.vram_model_reserved_bytes was being measured on a card without the model. The attest pod is a bare gt-attest:v1 card, so the checker reported the attest server's own 573 MB; in the release that makes attest.py expect ~32 GB free on a 5090, and a card holding the 23.7 GB model can fill only ~8 GB < 0.6 × 32 GB → every honest miner under-filled. The checker now reports the field only from a card with ≥ 8 GB resident and the pin-bump keeps the loadout's existing value otherwise (jq fallback verified).

Pin-bump now carries aggregate_decode_tps into the loadout — #1737 made the checker emit it and derives the per-token rate from it, but the jq step only copied single_stream_decode_tps and decode_per_request.

A non-JSON lium ls line killed run 33529543919 before renting (jq: parse error … column 6 — the shape of a leading Hint:/Error: line). free_node now drops anything before the first [, checks the listing parses as an array, and otherwise logs and retries.

ruff / format / pyright clean on the checker; bash -n, shellcheck -S warning, YAML parse clean.

https://claude.ai/code/session_011rz9LRxHEUXn37Jjqbxxh9