Nice fix — the routing + drain mechanics are clean, the comments are excellent, and the node_test_module level 0a case (concurrency + mixed t.test()/routed test()) is good coverage. overrides pass-through and the defensive rejection handler on PromisePrototypeThen both look right. CI is green so far.

One concern I want on the record before this is relied on for messy suites: the module-global currentTestContext with synchronous save/restore is only correct for stack-disciplined (LIFO) execution, but the unawaited-routed-subtest path this PR enables is exactly where execution stops being LIFO. A routed subtest with an await in its body can leave currentTestContext pointing at a sibling/child when the parent resumes:

test("parent", async (t) => {                       // current = parent
  test("childA", async () => { await delay(50); });  // routed; childA suspends with current = childA
  await delay(10);                                   // parent suspends
  // childA's body resumes first: current = childA, then awaits delay(50)
  test("childB", () => {});                          // parent resumes while current === childA -> mis-routed under childA
});

childB registers as a subtest of childA instead of parent — silent mis-nesting (and if the wrong context's step has already finished, this.#denoContext.step(...) throws instead). It needs a routed subtest whose body awaits plus a parent that awaits and then issues another routed test(), which is plausible in fastify-style suites.

The 8330212 commit message explains the als.run(...) revert: it added stack frames that pushed ext:cli/40_test.js out of the failure stack and broke the LSP source location. That is a real constraint — but AsyncLocalStorage.enterWith(nodeTestContext) avoids it: it's a synchronous setter that adds no stack frame, yet still propagates across await and into child async contexts, so each interleaved body sees its own context without manual save/restore. Each Deno test/step fn is its own async root, so the store stays scoped to that body's subtree.

Not blocking from my side (your PR, your call) — but worth either confirming the interleaving case behaves or switching to enterWith.