State meta-learner assumptions and correct X-learner weighting

[AliFozooni]

Summary

• state the consistency, conditional-exchangeability, and overlap assumptions needed to train causal meta-learners on observational data
• distinguish randomized validation from identification in the observational training sample
• separate overlap diagnostics from the transport assumptions needed for a new target population
• define both X-learner imputed-effect targets and second-stage models explicitly
• correct the propensity-weighting explanation by tracking which counterfactual outcome model supports each effect estimate
• clarify that propensity blending is a practical choice, not a repair for hidden confounding or non-overlap

Why this matters

Flexible outcome models do not make an observational comparison causal by themselves, and a randomized test set cannot remove confounding learned during training. The revised introduction makes those identification limits explicit.

The X-learner weighting can also look reversed if it is explained only by the arm used in the second stage. When treated observations are locally abundant, the treated-outcome model has better support, so e(x) weights the control-unit effect model that uses it to impute the missing treated outcome. When controls are abundant, the complementary weight favors the treated-unit effect model that uses the better-supported control-outcome model.

Validation

• built the full Jupyter Book successfully
• inspected the rendered assumptions and weighting equations
• validated notebook JSON and schema
• confirmed that only three markdown cells changed and all code, outputs, and execution counts are unchanged

Closes #398
Closes #454