docs(ml): clarify edge-case comments from final review

- ab_testing: clarify champion_dd == 0 drawdown semantics
- online_learning: document compute_fisher independent-loss requirement
- curriculum: document disabled-vs-override precedence

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

crates/ml/src/deployment/ab_testing.rs

@@ -642,8 +642,9 @@ impl PromotionCriteria {
             };
         }
 
-        // 3. Drawdown constraint: challenger DD must not exceed ratio * champion DD
-        //    Guard against champion_dd == 0: any positive challenger_dd would fail.
+        // 3. Drawdown constraint: challenger DD must not exceed ratio * champion DD.
+        //    When champion_dd == 0, dd_limit == 0 so any positive challenger_dd fails
+        //    (correctly strict). Both == 0 passes (both perfect — valid).
         let dd_limit = champion_dd * self.max_drawdown_ratio;
         if challenger_dd > dd_limit {
             return PromotionDecision::NotReady {

crates/ml/src/trainers/curriculum.rs

@@ -76,8 +76,8 @@ pub struct CurriculumScheduler {
 impl CurriculumScheduler {
     /// Create a new scheduler.
     ///
-    /// If `config.override_phase` is set, the scheduler starts in that phase.
-    /// If curriculum learning is disabled, the scheduler starts in `AllRegimes`.
+    /// Precedence: `!enabled` (→ AllRegimes) > `override_phase` > default (Basic).
+    /// When disabled, `override_phase` is ignored — all actions/regimes are allowed.
     pub fn new(config: CurriculumConfig) -> Self {
         let initial_phase = if !config.enabled {
             CurriculumPhase::AllRegimes

crates/ml/src/trainers/online_learning.rs

@@ -183,6 +183,10 @@ impl EWCRegularizer {
     /// accumulate the squared gradient magnitudes (diagonal Fisher approx):
     ///
     ///   F_i = (1/N) * sum_n (d loss_n / d theta_i)^2
+    ///
+    /// **Important**: each element of `losses` must come from an independent
+    /// forward pass. Calling `backward()` consumes gradients, so reusing a
+    /// loss tensor that has already been back-propagated produces zero grads.
     pub fn compute_fisher(
         &mut self,
         var_map: &VarMap,