perf(ssr): optimize isCircularImport algorithm for large codebases

[hanford]

Description

I swapped out the old recursive “visited” set for a simple 3-state map (unvisited, visiting, done) that walks each module only once. As soon as we hit a module marked “visiting,” we know there’s a cycle and bail out.

Why It Matters

1. Old code would re-scan big dependency trees many times
2. The single integer map tracks both “on the stack” and “already safe” status
3. Early exit on cycle means less wasted time/work
4. Startup time dropped noticeably in our big repo

fixes #20084

[sapphi-red]

Thank you! This is nice.

Startup time dropped noticeably in our big repo

Do you have a concrete value of this? I understand that the performance should improve theoretically due to the reduced algorithmic complexity, but I’m curious about the actual numbers.

[hanford]

I can share concrete numbers tomorrow morning!

[hanford]

I think I found some issues with this approach, will report back after taking a deeper look

see comment above

[sapphi-red]

I'll convert this to draft for now.