Clarify full-sample fitting under the conformal sharp null

[AliFozooni]

Summary

• explain why the counterfactual model is intentionally refit on the complete null-adjusted series
• connect the implementation directly to sharp-null imputation in Chernozhukov, Wüthrich, and Zhu
• explain how the complete-series fit puts the designated post-treatment block and cyclic reassignments on the same footing
• clarify why a pre-treatment-only fit would mix in-sample and out-of-sample residuals and define a different calibration
• remove the misleading suggestion that full-sample fitting merely prevents large residuals
• correct the p-value prose so it matches the displayed indicator and code: count permuted statistics at least as large as the observed statistic

Why this matters

Under a candidate sharp null, subtracting the candidate effect imputes the missing untreated outcome in every post-treatment period. The cited procedure then estimates the counterfactual model on that complete candidate path and evaluates the designated post-treatment residual block against cyclic block reassignments. The existing code follows that published algorithm; changing it to fit only before treatment would not be a bug fix, but a different procedure with a fitting-status break at the intervention date.

The p-value is the upper-tail permutation proportion. The formula and implementation already use the correct direction, but the surrounding sentence described the reverse inequality; the text now agrees with both.

Validation

• checked the explanation against the primary JASA paper and retained its DOI link
• built the full Jupyter Book successfully after both corrections
• inspected the rendered HTML, equations, and citation link
• validated notebook JSON and confirmed that only two markdown cells changed

Closes #344