[miguel-heygen]

What changes

The edit accuracy shards only need packages/cli/dist, but they waited for the whole Build job, which goes on to pack every package (verify:packed-manifests) and check the CLI's lazy dependencies after the CLI already exists. A new job, Studio: edit accuracy CLI build, runs the same install and bun run build and uploads the CLI; the shards and their retry job need it instead of Build.

Build keeps every step and stays the required check. It no longer runs just because the edit accuracy filter matched, since that clause (from #5566) only existed to produce the artifact.

If the new job fails, the shards are skipped and Studio: edit accuracy gate fails because the shards did not run (#6069), so a broken CLI build still blocks the merge.

Why

The shards are the merge queue's critical path: Detect changes -> Build -> the slowest of 20 shards -> the gate.

Before, over the last 20 green merge-queue CI runs: Build took a median 266 s, of which verify:packed-manifests took a median 98 s and check:cli-lazy-deps 4 s, all after the CLI was built. In run 40985995480 the CLI build finished 114 s into Build; the shards started 202 s in.

After

This PR's own run (40989910509, pull request, 5 shards): Studio: edit accuracy CLI build ended 195 s into the run and the shards started at 197-198 s; Build ended at 285 s. Before this change the shards could only start after Build, so they now start about 87 s sooner, and the run's critical path is shorter by the same amount. Merge-queue numbers follow from the first queue runs after it lands.

Checks

• node --import tsx --test scripts/check-edit-accuracy-filter.test.mjs scripts/edit-accuracy-retry.test.mjs scripts/merge-queue-workflows.test.mjs scripts/ci-cache-workflows.test.mjs: 9/9 pass. The filter test now reads the new job's steps for what the bench is built from.

Size

Under 100 lines on purpose: one workflow dependency change that takes 87 s off every CI run; nothing else open touches ci.yml to carry it.

[github-actions]

Edit accuracy: accurate 271 (base branch 271), smooth 167 of those

The gate passes.
Smoothness is reported in the artifact, not gated. A case fails only if it fails 2 of 3 runs.

Quarantined, measured but not gated (0)

[heygengenesis]

Verdict: APPROVE · 0 blocker · 0 important · 0 nits · head 0c89941a

Scope

Read the ci.yml diff in context. The only thing that consumed the edit-accuracy-cli artifact was the shards and their retry, and both now need the new job. Dropping edit_accuracy from Build's if only skips Build on studio-only diffs, and a skipped Build reads as passing. It was only there to produce the artifact. Fail-safe holds: if the CLI build fails, the shards are skipped and ratchet.mjs gate --shards returns "0 of N shards reported results", so the required gate fails. The 4 workflow test files pass locally (9/9). Both CI runs on this head are green.