TrackedContextualBandit

Observability wrapper around any ContextualBandit. Every event flows into a small set of aggregate side stats, each modelling a different question about the bandit's behaviour. Arm-level bucketing is a separate stratify op; until that lands, encode the arm into the observation (via the joint template) and read the slope/contrast off the resulting stat.

Templates are independent and any subset may be null:

• chooseTemplate sees update(x = context, y = armIndex.toDouble(), weight = 1.0) at every choose. Models the bandit's policy; the distribution of arm selections as a function of context.

• updateJointTemplate sees update(x = [armIndex.toDouble()] ++ context, y = reward, weight) at every update. Joint reward model with the chosen arm prepended as an extra feature; the coefficient on the arm dimension is the arm-conditional effect. featureSize of this template must equal 1 + contextFeatureSize.

• updateMarginalTemplate sees update(x = context, y = reward, weight) at every update. Marginal reward-given-context model, agnostic to arm. featureSize must equal contextFeatureSize.

• updateArmRewardTemplate sees update(x = armIndex.toDouble(), y = reward, weight) at every update. Per-arm reward distribution expressed as a paired stat; covariance, correlation, or per-arm slope.

The wrapper itself only satisfies ContextualBandit; the underlying bandit is exposed as inner typed B so callers reach extra interfaces; snapshot(), armResult, evaluate(i, x); through tracked.inner.<method> without losing static type information.