Health/probe gap: /api/v1/health can't distinguish two origins with different data; EU served an empty /api/v1/scores for 7 months

[emooreatx]

What went unnoticed

api.ethicsengine.org has two A records (US 207.148.13.157, EU 91.99.150.121), unproxied. Both returned identical /api/v1/health (version 1, build 45, "CIRISNode is healthy") for seven months while:

	US	EU
/api/v1/scores	8 models	{"scores": []}
/api/v1/embed/scores	1534 B	119 B
evals per model	5–6	2–4 (below the 5-eval publish threshold)

Root cause is infrastructure (databases never replicated — CIRISAI/CIRISCore#2), but the application gave a probe nothing to catch it with, and from North America happy-eyeballs picked the US origin ~100% of the time so manual checks looked fine.

Asks

1. Data-aware health: add /api/v1/health/data (or extend /api/v1/health) with {"evaluations": N, "publishable_models": M, "latest_completed_at": ts, "db_host_fingerprint": <hash of current_setting('cluster_name')||inet_server_addr()>} so a per-origin probe can diff two origins in one line.
2. /api/v1/health should report the real build: git SHA / image tag, not a hard-coded version: 1, build: 45. Every origin has said "build 45" through at least five deploys.
3. Ship a tiny per-origin probe (script or GitHub Action) that --resolves each A record and alerts when publishable_models differ. It belongs in CIRISCore's ops alerting, but the endpoint has to exist here first.

Found during the 2026-09-24 site audit.