fix(server): log why a readiness probe failed instead of discarding the cause

[ntatschner]

Production is currently returning 503 from /readyz with minio: "fail" — and there is no way to find out why.

{"status":"degraded","version":"1.8.132","checks":{"postgres":"ok","spicedb":"ok","minio":"fail"}}

The gap

ready() collapsed each probe with .is_ok(), discarding the error:

let minio_status = match minio.as_ref() {
    Some(mirror) => Some(mirror.ping().await.is_ok()),   // <- cause dropped here
    None => None,
};

The probes build genuinely useful errors. MinioMirror::ping() is a HeadBucket wrapped with:

MinIO HeadBucket failed for '{bucket}' (bucket missing or credentials invalid)

None of that is logged, and the response vocabulary is one word. So an unreachable endpoint, a missing bucket, and a rotated credential are indistinguishable — from outside and from the logs. That is the state right now: the deployment is degraded and nothing recorded which of the three it is.

Change

Each probe logs its full cause chain before being collapsed to the bool the response needs. Postgres and SpiceDB get the same treatment — they have the identical blind spot, and a silently unexplained Postgres readiness failure would be worse than this one.

WARN readyz: minio probe failed  error="MinIO HeadBucket failed for `starstats-audit` (bucket missing or credentials invalid): dispatch failure: io error: ..."

Uses {e:#} (single-line, JSON-pipeline friendly) — consistent with smtp_admin_routes.rs:351 and #331.

Logged server-side only — deliberately

The cause is not added to the response body. /readyz is publicly reachable (I read the 503 above from the open internet) and the text carries bucket names and endpoints. The body keeps its coarse ok/fail/skipped/degraded vocabulary — no schema change, no new information disclosure.

Note on scope

This does not fix MinIO. It makes the next probe say what is wrong, which is the blocker for fixing it. Given the mirror is best-effort (Postgres is the audit source of truth and writes still succeed), the practical impact today is that every mirror PUT is failing silently — the /readyz 503 is the only signal, and it is unexplained.

cargo test -p starstats-server 959 + 931, clippy -D warnings + cargo fmt --all --check clean. No migration, no behaviour change beyond log output.