Analysis threads need big stacks: deep type-inference recursion on
pathological files overflows rayon’s 2 MB default, and salsa 0.28’s
DependencyGraph::update_transferred_edges recurses per transferred
dependent under contended parallel queries — observed blowing 16 MB on
the Laravel fixture (64 MB measured clean over repeated runs; the pages
are reserved, not committed). Size the global pool before any parallel
analysis runs; build_global fails harmlessly if a pool already exists.
Benches/tests driving AnalysisSession without a DocumentStore must
call this themselves — rayon otherwise lazily builds the pool with
default stacks on mir’s first par_iter.