D785 — CI splits its slow legs across runners

Decision. The Windows tests run in four shards on separate runners (scripts/test-shard.sh), with install.ps1 on a fifth, and the browser suite runs in two legs: the @video-tagged export test, which also carries the frontend unit tests and bundle checks, and the rest. A small job named test-windows and one named web collect the legs and are the required checks. On main the concurrency group is the commit, so every merge gets its own run.

Why. One Windows job ran make gate for 12 minutes on average and up to 19, almost all of it waiting on internal/mcpserver (about 13 minutes, against one on ubuntu: git process creation is that much slower on Windows), close to the package's 20-minute test timeout. One browser test, the replay video export, was half of the five-minute web job. Standard runner minutes cost nothing on a public repository, so wall time is the only cost that matters. The price is more moving parts: a sharding script and two collector jobs.

Alternatives rejected. - A bigger Windows runner: larger runners are billed even on public repositories. - Sharding by package: internal/mcpserver alone would still be the whole wait. - One test-name regex for the whole repository: past Windows' command-line limit. - Playwright --shard: with fullyParallel: false it splits by file, and the slow test's file would still be the long leg; a tag says which test is slow. - Matrix legs as the required checks: a matrix renames them, which silently un-protects main.

Consequences. The Windows shards run the tests, not make gate: fmt-check and vet (with GOOS=windows) stay on test, and internal/repodocs runs only there. A new test lands in a shard by itself; a new slow browser test needs the @video tag or a leg of its own. Branch protection must keep requiring the collector names, never a leg's.