A cached vp run spends seconds compiling the same cache globs while building the task graph #886

Written by an AI agent (Claude) on behalf of @jasonkuhrt.

In a workspace of 138 packages and 1,576 tasks, vp run @libs/enum#test with both of its tasks cached spends about 2 s of CPU after the configs load and before the first task starts, building the task graph.

Cause. resolve_glob_to_workspace_relative calls wax::Glob::new for every input and output pattern of every task. In wax 0.7, Glob::new compiles a regular expression and partition compiles another, and only the partition's prefix and text are kept. The workspace has 13,708 pattern uses but only 1,011 distinct patterns: packages repeat the same src/** and the same negations.

Measured (Vite+ 1.4.0 built against vite-task a3c087a, optimized binding, macOS, 3 interleaved runs per variant, load average about 7):

A second, separate idea. Configs load one package after another (lib.rs). A prototype that let up to 8 loads overlap, keeping results in order, returned the same 138 configs as the serial load and took the same run from 2.47 s to 2.19 s CPU (1.79 s to 1.62 s wall) on top of the glob change. A config whose result depends on load order would be the risk, so it is left out of the fix.