test(flake): per-test event-loop yield in the mocha root hook, scoped to etherpad's own backend specs

PR #7854's first iteration added the yield to 6 known-dying spec files
(pad.ts, importexportGetPost.ts, socketio.ts, messages.ts, import.ts,
clientvar_rev_consistency.ts). Linux backend matrix passed, proving
the yield doesn't break the affected tests' own state-sharing
assumptions. But the very next Win+plugins run captured **death #13
in sessionsAndGroups.ts**, a 7th file outside the scoped fix. The
flake migrated rather than being suppressed.

That's strong evidence the trigger is the rapid-sequential-test
pattern in general, not specific files. Replace the per-file scope
with a root-level `mochaHooks.beforeEach` yield in diagnostics.ts,
gated on a file-path check: yield for ether/etherpad's own specs in
`tests/backend/specs/`, SKIP for plugin tests loaded from
`../node_modules/ep_*/static/tests/backend/specs/`.

The plugin-test skip exists because PR #7844 demonstrated that an
unconditional global yield breaks `ep_subscript_and_superscript`'s
`returns HTML with Subscript HTML tags` series — those plugin tests
share state across describe-block boundaries and don't tolerate any
microtask reordering. The file-path check preserves PR #7844's
finding without re-breaking those tests.

Files modified:
  - src/tests/backend/diagnostics.ts: root beforeEach yield, scoped

Per-file changes from the previous commit are reverted — root scope
supersedes them and there's no point yielding twice per test.

Test plan unchanged from the original PR:
  - Linux ± plugins must pass.
  - Windows ± plugins flake rate: ~22% pre-fix. Post-fix, run the CI
    5-10x and compare. If unchanged, cadence is ruled out as the
    trigger and we look at per-test pathologies (jose CNG on
    Windows, libuv IOCP edge cases unrelated to load).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: JohnMcLear

src/tests/backend/diagnostics.ts

@@ -210,10 +210,28 @@ for (const sig of ['SIGINT', 'SIGTERM', 'SIGHUP', 'SIGBREAK'] as const) {
 // resolution fails us. A `start` line per test gives sub-millisecond
 // resolution on which test was on the rails when the process died.
 export const mochaHooks = {
-  beforeEach(this: any) {
+  async beforeEach(this: any) {
     if (this.currentTest) {
       currentTest = this.currentTest.fullTitle();
       diag(`test start: ${currentTest}`);
+      // Per-test event-loop yield for ether/etherpad's own backend tests
+      // (not plugin tests). Mitigation against the Windows silent-ELIFECYCLE
+      // flake: 13 captures so far show V8 starvation 200-400 ms before each
+      // kill in test files that fire 50-100 short tests in sequence. The
+      // initial scoped fix in PR #7854 covered 6 known-dying files but the
+      // flake migrated to a 7th (sessionsAndGroups.ts) on the next run —
+      // so the trigger is the broader rapid-sequential-test pattern, not
+      // specific files. A root-level yield covers all our backend tests.
+      //
+      // Plugin tests (loaded from `../node_modules/ep_*/static/tests/
+      // backend/specs`) are SKIPPED via file-path check because PR #7844
+      // demonstrated the global variant breaks ep_subscript_and_superscript
+      // tests that share state across describe-block boundaries. The check
+      // is conservative: only ether/etherpad's own specs get the yield.
+      const file = this.currentTest.file as string | undefined;
+      if (file && !/node_modules[/\\]ep_/.test(file)) {
+        await new Promise((r) => setImmediate(r));
+      }
       // Drop a node-report at test-boundary granularity when the inter-report
       // gap is wide enough. Run 26399285213's rerun caught the kill on the
       // socketio.ts duplicate-author test, but the previous boundary write