\section{Robustness}
\label{sec:robustness}

The Internet Appendix reports the full battery specified in the analysis
protocol. The main text highlights the checks that discipline
interpretation. \emph{Tick-regime}: pre-2001 minute returns embed
eighth/sixteenth-tick bid--ask bounce; all secular comparisons are re-run on
range-based measures, on 5- and 15-minute aggregation, and within
post-decimalization subsamples. \emph{Coverage}: the 2020 source-coverage
expansion is absorbed by fixed-rule lagged top-$N$ universes; results are
reported for $N\in\{500,1000,3000\}$. \emph{Universe composition}: ETFs and
likely non-common instruments (ticker-shape heuristic) are excluded in
robustness runs. \emph{Corporate actions}: split windows are excluded
throughout; sensitivity to window width and detection thresholds is
tabulated. \emph{Thresholds}: source shocks at $|z|\ge4,5,6$ and receiver
thresholds at $|z|>2,3$ leave the baseline null unchanged. \emph{Inference}:
day-block and month-block bootstrap agree. Table~\ref{tab:robustness}
collects the headline checks: the market-wide null and the $+64\times10^{-4}$
network differential are stable across the post-decimalization subsample,
mid-day-only events, announcement-window exclusions, calm and crisis years,
higher source thresholds, and non-common-ticker exclusions, and the neighbor
differential measured on placebo events alone is zero ($-2.7\pm2.8\times
10^{-4}$)---the permuted-benchmark falsification the design commits to. The
one systematic pattern is that the network differential \emph{declines} with
source magnitude ($+64$ at $|z|\ge6$ to $+34\times10^{-4}$ at $|z|\ge12$),
consistent with the largest events being firm-specific news that neighbors
correctly do not price.

\input{tables/table8_robustness}