Distinguish mediator adjustment from collider bias

[AliFozooni]

Summary

• distinguish conditioning on a mediator from conditioning on a collider
• explain total effects and controlled direct effects with explicit potential-outcome estimands
• state why randomized treatment alone does not identify a direct effect after conditioning on an observed mediator
• show the additional graph under which a mediator can also act as a collider
• clarify that blocking an indirect path need not always attenuate the estimate
• add explaining-away terminology to the earlier collider example

Why this matters

The previous discussion treated the white-collar employment mediator as though conditioning on it necessarily opened a collider path. In the displayed graph, it instead blocks the mediated causal path and changes the estimand from a total effect to a direct effect. Collider bias arises only after adding another cause of both the mediator and outcome. Keeping these mechanisms separate is essential for choosing an adjustment set and interpreting the resulting estimate.

Validation

• built the full Jupyter Book successfully
• validated the notebook JSON and checked the diff for whitespace errors
• confirmed that only four markdown cells changed and no stored outputs changed

Closes #424
Closes #462
Closes #467