Correct LATE indexing and the Chapter 9 OLS benchmark

[AliFozooni]

Summary

• label the randomized-assignment contrast as ITT, the receipt comparison as a naive association, and the Wald estimand as LATE
• explain why noncompliance dilutes ITT without making ITT biased for the assignment effect
• state instrument independence jointly over the complete potential-outcome and potential-treatment vector used by the Wald proof
• correct the double-index potential-outcome description: the first index is treatment and the second is instrument assignment
• make the naive OLS code match the delivered-versus-undelivered contrast defined in the text
• regenerate the OLS table and update the comparison from 27.60 to 13.93
• explain why IV2SLS.from_formula without a bracketed IV block is OLS

Why this matters

The notebook previously used ATE for three different quantities: the randomized effect of assignment, a confounded comparison by treatment receipt, and the Wald ratio. They are not the same estimand. Random assignment identifies the ITT; the observed receipt contrast is noncausal here; and under joint independence, exclusion, a first stage, and monotonicity, the Wald ratio identifies the average treatment effect for compliers, or LATE, rather than the population ATE. Monotonicity rules out defiers, while the instrument leaves treatment unchanged for always- and never-takers.

The notebook also defines its naive benchmark as E[Y | push delivered = 1] - E[Y | push delivered = 0], but the code additionally conditioned on push assignment. Because assignment is the instrument, that regression answered a different question and produced the 27.60 coefficient used by the prose. The corrected binary-treatment regression is exactly the observed delivered-versus-undelivered difference in means: 13.931. The 10.159 value discussed in #393 compares selected assignment and treatment cells rather than the unadjusted receipt groups.

Validation

• executed the modified OLS cell against the included app-engagement data
• regenerated the parameter table and synchronized its execution count
• built the complete Jupyter Book successfully after all estimand and assumption corrections
• validated notebook JSON and checked the diff for whitespace errors

The full build retains one unrelated pre-existing Chapter 21 lexer warning.

Closes #351
Closes #381
Closes #387
Closes #393