fix(outcome): forgive candidate regressions when clears outnumber them

src/brigade/outcome.py

@@ -400,6 +400,24 @@ def split_by_capability(
     return CapabilityCohorts(current_capability, pooled, capability, shrunk, off_capability, capability_legacy)
 
 
+def candidate_regression_withheld(helped: int, hurt: int) -> bool:
+    """Return True when a candidate cohort should stay withheld for regressions.
+
+    One verified hurt used to withhold the unchanged content-fingerprint cohort
+    forever. Later verified clears in that same cohort can release the hold once
+    they strictly outnumber hurts. Append-only evidence is never rewritten;
+    demotion of an already promoted artifact stays a separate policy.
+    """
+    return hurt > 0 and helped <= hurt
+
+
+def candidate_install_reason(helped: int, hurt: int) -> str:
+    """Reason string for a candidate install that cleared the hurt gate."""
+    if hurt > 0:
+        return "verified helped outnumbers regressions"
+    return "verified helped, no regressions"
+
+
 def decide(
     score: OutcomeScore,
     *,
@@ -418,10 +436,15 @@ def decide(
         return Decision(score.artifact_id, "hold", current_status, "cooldown active")
 
     if current_status == "candidate":
-        if score.hurt > 0:
+        if candidate_regression_withheld(score.helped, score.hurt):
             return Decision(score.artifact_id, "hold", "candidate", "withheld: verified regression present")
-        if score.helped >= config.install_min_helped:
-            return Decision(score.artifact_id, "install", "promoted", "verified helped, no regressions")
+        if score.helped >= config.install_min_helped and score.helped > score.hurt:
+            return Decision(
+                score.artifact_id,
+                "install",
+                "promoted",
+                candidate_install_reason(score.helped, score.hurt),
+            )
         return Decision(score.artifact_id, "hold", "candidate", "insufficient verified evidence")
 
     if current_status == "promoted":

src/brigade/scorecard.py

@@ -768,9 +768,9 @@ def effectiveness_gate_reason(
     hurt = int(effectiveness.get("hurt", 0))
     trials = helped + hurt
     wilson = outcome_core.wilson_lower_bound(helped, trials, config.z)
-    if hurt > 0:
+    if outcome_core.candidate_regression_withheld(helped, hurt):
         return "withheld: verified regression present"
-    if helped < config.install_min_helped:
+    if helped < config.install_min_helped or helped <= hurt:
         return "insufficient verified evidence"
     wilson_min = config.effective_wilson_min
     if wilson < wilson_min:
@@ -790,7 +790,10 @@ def dual_gate_passes(
     utility_reason = utility_gate_reason(card, min_passing_units=config.utility_min_passing_units)
     if utility_reason is not None:
         return False, utility_reason
-    return True, "verified helped, no regressions"
+    effectiveness = card.dimensions.get("effectiveness", {})
+    helped = int(effectiveness.get("helped", 0))
+    hurt = int(effectiveness.get("hurt", 0))
+    return True, outcome_core.candidate_install_reason(helped, hurt)
 
 
 def decide_scorecard(
@@ -826,7 +829,7 @@ def decide_scorecard(
         return outcome_core.Decision(artifact_id, "hold", current_status, "cooldown active")
 
     if current_status == "candidate":
-        if hurt > 0:
+        if outcome_core.candidate_regression_withheld(helped, hurt):
             return outcome_core.Decision(artifact_id, "hold", "candidate", "withheld: verified regression present")
         passes, reason = dual_gate_passes(card, config=config)
         if passes:

tests/test_outcome.py

@@ -115,16 +115,31 @@ def test_decide_installs_a_clean_candidate():
     assert decision.new_status == "promoted"
 
 
-def test_decide_withholds_candidate_with_a_verified_regression():
+def test_decide_withholds_candidate_when_regressions_are_not_outnumbered():
     decision = outcome.decide(
-        _score(3, 1),
+        _score(1, 1),
         current_status="candidate",
         last_action_ts=None,
         now=dt.datetime(2026, 6, 21),
         config=outcome.ReconcileConfig(),
     )
     assert decision.action == "hold"
     assert decision.new_status == "candidate"
+    assert decision.reason == "withheld: verified regression present"
+
+
+def test_decide_installs_candidate_when_later_clears_outnumber_regression():
+    """#629: helped=10 hurt=1 in one unchanged cohort must not withhold forever."""
+    decision = outcome.decide(
+        _score(10, 1),
+        current_status="candidate",
+        last_action_ts=None,
+        now=dt.datetime(2026, 6, 21),
+        config=outcome.ReconcileConfig(),
+    )
+    assert decision.action == "install"
+    assert decision.new_status == "promoted"
+    assert decision.reason == "verified helped outnumbers regressions"
 
 
 def test_decide_withholds_candidate_with_insufficient_evidence():
@@ -180,6 +195,39 @@ def _fp_record(evidence_ref, ts, fingerprint, signal=1):
     )
 
 
+def test_decide_scores_new_fingerprint_cohort_separately_from_regressed_cohort():
+    """A content-fingerprint change starts a new cohort; old hurts do not follow."""
+    records = [
+        _fp_record("r1", "2026-07-01T00:00:00+00:00", "old-rev"),
+        _fp_record("r2", "2026-07-02T00:00:00+00:00", "old-rev", signal=-1),
+        _fp_record("r3", "2026-07-03T00:00:00+00:00", "new-rev"),
+        _fp_record("r4", "2026-07-04T00:00:00+00:00", "new-rev"),
+    ]
+    cohorts = outcome.split_by_fingerprint("skill-x", records, "new-rev")
+    assert (cohorts.current.helped, cohorts.current.hurt) == (2, 0)
+    assert (cohorts.lifetime.helped, cohorts.lifetime.hurt) == (3, 1)
+
+    current_decision = outcome.decide(
+        cohorts.current,
+        current_status="candidate",
+        last_action_ts=None,
+        now=dt.datetime(2026, 7, 5),
+        config=outcome.ReconcileConfig(),
+    )
+    assert current_decision.action == "install"
+    assert current_decision.reason == "verified helped, no regressions"
+
+    lifetime_decision = outcome.decide(
+        cohorts.lifetime,
+        current_status="candidate",
+        last_action_ts=None,
+        now=dt.datetime(2026, 7, 5),
+        config=outcome.ReconcileConfig(),
+    )
+    assert lifetime_decision.action == "install"
+    assert lifetime_decision.reason == "verified helped outnumbers regressions"
+
+
 def test_split_by_fingerprint_grandfathers_legacy_and_drops_proven_stale():
     records = [
         _fp_record("r1", "2026-07-01T00:00:00+00:00", "old-rev"),