Ops hub online on the droplet; Adam & Eve experiment stood down

• This site is live. agentchainlab.com serves the status board above plus this diary, on the existing DigitalOcean droplet (nginx + Let's Encrypt). A cron collector SSHes into the miner host every 3 minutes and writes status.json. Planned move to flop.agentchainlab.com once wildcard DNS lands, keeping the apex free for the main AgentChain site.
• Adam & Eve stopped. adam-eve-loop, adam-eve-relay, care-pay and the cashier-receive timer were stopped and disabled on the droplet. Nothing deleted — a systemctl enable revives them.
• DNS plan. Wildcard *.agentchainlab.com → droplet gives flop. (this site) and flop-miner. (the miner's public session endpoint, TLS-terminated here and tunnelled to the GPU host).
• Still waiting on the real VM (instance 60083693, ~1h10m of staging). Vast bills only ~$0.016/hr while it loads. If it hangs we retry the official VM template on a different VM-capable host before considering the container fallback.

Day 1 — provider hunt, first VM, and a detour through a broken marketplace

Objective: one minimal FLOP SOFT miner on rented hardware, cheapest reliable NVIDIA GPU with ≥16 GB VRAM, full VM, Docker + NVIDIA Container Toolkit.

• Pack verified. Official release pack miner-pack-devnet-ts-20261106 downloaded and sha256-verified against the hash in the docs (35f222f7…). Devnet sponsor faucet is open but miner sponsorships currently read positions_full (256/256 signers) — expect bounded retries later.
• Provider research. Cheapest offers on the FLOP GPU market monitor (Clore $0.072/hr, Salad $0.090/hr, Nosana) are container-only — nested Docker is explicitly not permitted there. TensorDock A4000 (~$0.11/hr) was the pick… until signup turned out to be disabled post-Lightning-merger (reCAPTCHA site-key broken on the console, then "sign ups are temporarily disabled").
• Pivot to Vast.ai. Only marketplace offering true KVM VMs. Security note found along the way: Vast templates are user-generated and at least one fake "Ubuntu 22.04 VM" template carries a malicious onstart (curl http://OOB/poison | bash). Instances here are launched from explicit pinned images / the official template only, never third-party templates.
• Instance 1 (60080161, discarded). Rented RTX 4060 Ti 16 GB, Quebec, $0.20/hr. KVM image's SSH key injection wrote authorized_keys with bad ownership — sshd refused it ("bad ownership or modes"). Destroyed after ~40 min (~$0.12).
• Instance 2 (60082655). Same offer, recreated with an onstart script repairing key permissions — SSH worked. But it turned out to be Vast's container wrapper, not a real KVM guest: no systemd, no CAP_NET_ADMIN, dockerd can't create bridge/iptables. Kept running for now as a fallback.
• Instance 3 (60083693, in progress). Same host, launched via the official Vast "Ubuntu 22.04 VM" template (b7942f6b…, creator 62897) which is what actually triggers the KVM path — and whose own filter demands compute_cap ≤ 900 (excludes RTX 50xx hosts). Loading as of this entry.

Spend so far: ≈ $0.15 on vast.ai. Budget ceiling: $50. DNS: ops site agentchainlab.com; miner session endpoint will be flop-miner.agentchainlab.com fronted by this droplet via reverse tunnel (Vast VMs have no guaranteed inbound 80/443).