The `playwright-windows` job sharded the whole CLI e2e suite across two
shards, but the dev-server suites in `e2e/playwright/dev` dominate
Windows runtime. Because Playwright keeps `describe.serial` blocks on a
single shard, count-based sharding of the mixed suite left one shard at
~20m while the other finished in ~5m.

Split the slow `dev` directory across three `--shard` jobs and run the
faster `prod` suite on its own, reusing the same `--shard` mechanism as
the Ubuntu Playwright job. This drops the Windows Playwright test-step
long pole from ~20m to ~8m with the same total test coverage. Sharding
by directory keeps newly added test files covered automatically.