fix(jev): reject zero recoverability thresholds

Co-authored-by: Medulla <medulla@tinyhumans.ai>

Author: senamakel

crates/tinytools-jev/src/recovery/README.md

@@ -13,7 +13,10 @@ using evidenced corrections or transient conditions without an external
 prerequisite. Score is a three-category ordered feasibility distribution; its
 expectation (0..=2) is **not** a probability of success. Confidence is distribution
 concentration and is **not** accuracy. Hosts must calibrate explicit versioned
-thresholds, including the counted repeated-blocker threshold.
+thresholds, including the counted repeated-blocker threshold. Recoverability must
+be finite and in `(0, 1]`; class and advice concentration thresholds accept finite
+values in `[0, 1]`. A zero recoverability threshold would admit corrective advice
+with no prospect of recovery and reject valid external-prerequisite classifications.
 
 Candidate descriptions alone do not enable alternate evaluation. The host supplies
 `RecoveryAlternateReason::WrongToolAdvice` after prior advice or `RepeatedBlocker`

crates/tinytools-jev/src/recovery/validation.rs

@@ -18,13 +18,11 @@ pub(super) fn validate_thresholds(thresholds: &RecoveryThresholds) -> Result<(),
         true,
         RecoveryObservation::MAX_KEY_BYTES,
     ) || thresholds.repeated_blocker == 0
-        || ![
-            thresholds.class_confidence,
-            thresholds.recoverability,
-            thresholds.advice_confidence,
-        ]
-        .into_iter()
-        .all(probability)
+        || !probability(thresholds.recoverability)
+        || thresholds.recoverability <= 0.0
+        || ![thresholds.class_confidence, thresholds.advice_confidence]
+            .into_iter()
+            .all(probability)
     {
         return Err(RankError::invalid_input("invalid recovery thresholds"));
     }

crates/tinytools-jev/src/recovery/validation_tests.rs

@@ -95,6 +95,56 @@ async fn external_prerequisites_can_be_classified_without_corrective_advice() {
         }
     ));
 }
+#[test]
+fn recoverability_threshold_requires_a_finite_positive_probability() {
+    for value in [
+        0.0,
+        -0.0,
+        -0.1,
+        f64::NAN,
+        f64::INFINITY,
+        f64::NEG_INFINITY,
+        1.1,
+    ] {
+        let mut policy = thresholds();
+        policy.recoverability = value;
+        let fake = Arc::new(Fake {
+            calls: AtomicUsize::new(0),
+            result: Err(()),
+        });
+        assert!(
+            RecoveryAdviser::new(fake, policy).is_err(),
+            "invalid recoverability threshold {value}"
+        );
+    }
+    for value in [f64::MIN_POSITIVE, 1.0] {
+        let mut policy = thresholds();
+        policy.recoverability = value;
+        let fake = Arc::new(Fake {
+            calls: AtomicUsize::new(0),
+            result: Err(()),
+        });
+        assert!(
+            RecoveryAdviser::new(fake, policy).is_ok(),
+            "valid recoverability threshold {value}"
+        );
+    }
+}
+
+#[test]
+fn concentration_thresholds_can_include_zero_and_one() {
+    for value in [0.0, 1.0] {
+        let mut policy = thresholds();
+        policy.class_confidence = value;
+        policy.advice_confidence = value;
+        let fake = Arc::new(Fake {
+            calls: AtomicUsize::new(0),
+            result: Err(()),
+        });
+        assert!(RecoveryAdviser::new(fake, policy).is_ok());
+    }
+}
+
 #[test]
 fn thresholds_and_observation_boundaries_are_explicit() {
     for invalid in [f64::NAN, -0.1, 1.1] {