fix(e2e-mock): replace sequenceIndex with predicate match for abandon parent fixture

The abandon-subtask parent fixture used userMessage + sequenceIndex: 0 for
matching. sequenceIndex shares a global counter (DEFAULT_TEST_ID) across all
e2e tests, making it fragile against stream interleaving from prior tests'
delayed mock responses. When the counter was stale or a delayed stream from
the previous test interleaved, the wrong fixture (interrupt parent
completion) could match, returning attempt_completion with new_task-style
arguments, causing a tool-call parse failure and 30s timeout.

Replace with a predicate match using lastUserMessageContains +
SUBTASK_RESULT_INJECTION exclusion, mirroring the pattern used by the
regular and fast parent fixtures. This eliminates the sequenceIndex
dependency and ensures the fixture only matches the parent's initial
request, not resume turns.

Author: Zoo (VP)

apps/vscode-e2e/src/fixtures/subtasks.ts

@@ -565,10 +565,17 @@ export function addSubtaskFixtures(mock: InstanceType<typeof LLMock>) {
 	})
 
 	// Abandon-subtask scenario (#559)
+	// Use a predicate (lastUserMessageContains + SUBTASK_RESULT_INJECTION exclusion) instead of
+	// userMessage + sequenceIndex. sequenceIndex shares a global counter across all tests
+	// (DEFAULT_TEST_ID), making it fragile if a prior test's delayed stream interleaves with
+	// this fixture's match. The predicate approach mirrors the regular parent fixtures and
+	// ensures the initial new_task call is served only on the parent's first turn, not on a
+	// resume turn (which carries SUBTASK_RESULT_INJECTION from the delegated child result).
 	mock.addFixture({
 		match: {
-			userMessage: new RegExp(SUBTASK_ABANDON_PARENT_MARKER),
-			sequenceIndex: 0,
+			predicate: (req: ChatCompletionRequest) =>
+				lastUserMessageContains(req, SUBTASK_ABANDON_PARENT_MARKER) &&
+				!requestContains(req, [SUBTASK_RESULT_INJECTION]),
 		},
 		response: {
 			toolCalls: [