Required status-check names stay on every OS/Node cell so the main
ruleset still sees a report. pull_request cells just do less work.

- Drive scope from ts/tools/scripts/prCiScope.mjs (full / ratchet /
  package). push and merge_group still run the full matrix.
- build-ts PRs: skip Node 24 install/build/test; run ratchets and a
  single git fetch of the base on ubuntu/22 only.
- build-ts PRs: shallow checkout on non-ratchet cells.
- build-package-shell PRs: keep ubuntu packaging; skip macos/windows.
- azure-smoke-tests detect job: fetchDepth 2 (HEAD and HEAD^1).
- Add node --test coverage that invokes the real CLI and reads the
  shipped workflow YAML.

Skipping Node 24 tests or macos/windows packaging would make the PR
gate weaker than the merge. Restore that work.

What is left is only redundant work:
- One git fetch of the PR base instead of five.
- Ratchets once on ubuntu/22 (same tree, not OS-specific).
- Shallow checkout on cells that only install/build/test.
- ADO detect job fetchDepth 2 (smoke tests still run).

Every OS/Node cell still builds, tests, and packages.

- Checkout PRs at depth 2 (merge commit + parents) on all build-ts cells.
  Fetch the base ref once with --depth=1.
- ADO smoke: install Playwright in parallel with npm run build.
- ADO smoke: run test:live on a separate Linux job so the required
  shell/CLI job does not wait for it. Live tests still run
  (continueOnError unchanged).

Realtime scanning of checkout, pnpm, and test output is a known
multi-minute tax on GitHub-hosted and ADO Windows agents. The same
suites still run on every required cell.

Also record the #2847 required-check span (65.90 min) and the 46.13 min
30% target in the measurements note.

The required GitHub check "TypeAgent Smoke Tests" stays queued until
the ADO pipeline finishes. On SHA 82791dc the Windows leg was 33 min
because it ran the full Playwright+jest suite (shell:test) while Linux
only ran shell:smoke (simple.spec.ts).

Pull requests now run shell:smoke on both OS. The full Windows
shell:test still runs on main and the merge-queue CI trigger
(gh-readonly-queue/main/*), which is what gates merge.

Suite choice comes from prCiScope.mjs --windows-shell-suite.

- Skip live_linux on PullRequest. continueOnError already meant a live
  failure did not fail the gate, but the parent TypeAgent Smoke Tests
  check stayed queued until live finished (~13 min past Windows smoke on
  7e4135e). main and merge-queue still run test:live.
- Smoke agents build only agent-shell|agent-cli (with deps) and install
  Playwright chromium only. playwright.config.ts has one project;
  shell:smoke launches Electron. Full monorepo build stays on live and
  build_ts.
- On 7e4135e the required span was 36.78 min (ratio 0.558 vs #2847).
  Live was the pole; Windows smoke setup was 15.5 min of build+browsers
  for a 1.3 min suite.

Windows PR path no longer downgrades to shell:smoke. Full jest +
Playwright shell:test runs on PR, main, and merge-queue again. Linux
stays on shell:smoke (pre-existing).

30% bar still expected from the other cuts: skip PR live hold, ratchet
once, shallow fetch, Defender exclusion, shell+cli build scope,
Playwright chromium-only. Prior full-Windows parent was ~34 min vs
46 min target.

Remove workspace Defender disable/exclusion from build_ts, package
shell, and ADO smoke. The 30% PR wall-clock cut does not depend on it:
baseline already ran with default Defender; the gain is skip PR live
hold, single-cell ratchets, shallow PR fetch, and scoped smoke setup.