0694 — Alert evaluations return explicit states

Context

Evaluate returned an alert or nil, and Record treated nil as healthy. Pending and insufficient evidence must not resolve a recorded active alert. The user approved an explicit result provided every existing check uses it consistently.

Decision

• All three condition controllers implement Evaluator with AlertEvaluationResult: State is healthy, pending, active, or insufficient_evidence. Finding describes an unhealthy condition and is present only for pending and active results.
• Every scheduled worker keeps Evaluate -> Record. Record validates the result; pending and insufficient evidence return nothing without reading or writing the alert head. Healthy and active use the existing locked classification and production transaction. Only healthy can request recovery.
• Registration warnings read the same result and log findings for pending or active conditions. They remain immediate and log-only.
• These states are not new AlertStatus values or persisted evaluation state. Metrics retains ownership of evidence collection under [0693].

Consequences

The shared result is used immediately by every existing evaluator and caller. Missing partition measurements still return errors through existing check-failure reporting. History-derived duration, freshness, and pending config remain the next implementation work; this change does not enable them or change cadence. Alert-lab repair and execution remain deferred at the user’s request.