CLI: health at_risk includes deprecated artifacts (violates ADR-002 R_eff-skips-non-active)

[explosivebit]

Found in the v0.32.1 dogfood, re-verified explicitly against a workspace containing a deprecated artifact.

$ forgeplan health --json    (raw exit 0)
  → at_risk: [
      { id: "PROB-078", reason: "R_eff = 0.00 below 0.3 threshold",
        title: "MCP read-after-write returns stale body same session …" }
    ]
$ forgeplan get PROB-078 → Status: deprecated

PROB-078 is deprecated, yet it appears in the health at_risk signal because its R_eff is 0.00. ADR-002 states R_eff skips non-active artifacts (draft / deprecated / superseded); the at_risk aggregator should honor the same filter — a deprecated artifact at R_eff 0 is expected, not at-risk.

Related (same area): forgeplan score <deprecated-id> emits _next_action: null at R_eff=0 — a PRD-071 hint-contract gap; an at-risk score should still emit a Next hint.

Ask: filter at_risk to active-only artifacts in the health aggregator; emit a hint for R_eff=0 scores.

[explosivebit]

Confirmed still reproducing 2026-06-30 via MCP forgeplan_health.

Concrete instance: NOTE-026 (status=deprecated, phase=done, R_eff=0.00) — a deliberately-BIASED RIPER-E2E adversarial control whose zero R_eff is by design (EVID-171 correctly Refutes it; the C4 unbias gate working as intended). After clearing all 6 phase_mismatch anomalies in the workspace, verdict stayed needs_attention solely because this one deprecated artifact remains in at_risk.

Reinforces the ADR-002 'R_eff skips non-active' expectation: the at_risk set (and the verdict input) should exclude deprecated/superseded artifacts. Low-severity, but it produces a spurious needs_attention for agents / CI that branch on verdict, and sends cleanup loops chasing an intentionally-low-R_eff control.