Percentile cluster bootstrap under-covers below ~40 clusters #7

The implemented analysis is a plain percentile cluster bootstrap with no asymptotic refinement. Simulation of the exact procedure gives a realized rejection rate near 7% against a nominal 5% at 19 clusters with three replicates and realistic intra-case correlation, and the literature reports up to 10% in the 5-30 cluster range.

The limitation is now recorded in the plan and the release manifest (inference_limitations), which is the honest interim position, but a study at this scale should use a refinement.

References: Cameron, Gelbach & Miller (2008) on bootstrap-based improvements with clustered errors; Canay, Santos & Shaikh on the wild bootstrap with few large clusters; MacKinnon, Nielsen & Webb (2023).

Implementing this means changing analysis/paired-bootstrap.ts and analysis/system-bootstrap.ts and re-deriving the published interval semantics, so it needs its own change.