Various updates: move measurements fallback logic, add a delay between round-robin failover, update README & runbooks, make settings generic

[vreff]

This PR has the following thematic commits:

• Move fallback logic out of enclave client, moves the measurement fallback logic out of pool.go to keep that component simpler. Instead, its UpdateNodes function will return an error if the measurements aren't right, and the executor.go emits the same metric we alert on.
• Add delay between gateway round robin, adds a slight backoff period between our use of another gateway if one fails.
• Update README and release runbook, updates some naming in our runbook & readme
• Make settings more generic to the host, allow apps to parse them, we have a pretty app-specific settings object we pass through our whole framework. This change moves that to simply be bytes that can be passed as a JSON string from K8s and get parsed in the app. Once we start using this in infra, we will do a followup code change to move the type into the app and out of the global types.

[github-actions]

AI dashboard check

@vreff the AI Stupid Level dashboard is not currently showing any tracked models at score 55 or lower in the last 24 hours.

Tracked models currently not flagged

• GPT-5.4: gpt-5.4 (score 66, status good, trend stable)
• Claude Opus 4.5: claude-opus-4-5-20251101 (score 70, status good, trend up)
• Claude Opus 4.6: claude-opus-4-6 (score 65, status good, trend down)
• Claude Opus 4.7: claude-opus-4-7 (score 69, status good, trend stable)
• Claude Sonnet 4.5: claude-sonnet-4-5-20250929 (score 58, status warning, trend stable)
• Claude Sonnet 4.6: claude-sonnet-4-6 (score 65, status good, trend down)

To rerun this check on demand, comment /rerun-ai-dashboard-warning on this PR.

Snapshot: 2026-07-21T15:03:25.497Z from https://aistupidlevel.info (24h view).