test: widen timing margins on two Windows-flaky CI tests

Both failed on windows-latest today, on merges of otherwise-unrelated
PRs (#349, #356) - neither PR touched internal/specialist or
internal/agenteval. Windows GitHub Actions runners have coarser default
timer granularity and more scheduling contention than Linux/macOS, so
tight timing margins that hold on the other two platforms aren't
reliable there.

- TestOutputToolBlocksUntilTaskCompletes: a 100ms poll timeout against a
  5ms background write left only 20x margin. Widened the timeout to
  2000ms (a ceiling, not a target - PollInterval keeps the test fast in
  the success case; it only changes how much CI jitter it tolerates
  before failing).
- TestHarnessTimeoutCancelsBlockedAgent: confirmed there's no actual data
  race (harness.Agent.Run is a direct synchronous call, not a spawned
  goroutine racing the assertions) - the failure was workspace
  materialization alone consuming the full 10s budget on a contended
  Windows runner, so the agent was never reached. Widened to 30s. This
  test's own design blocks for the full timeout to observe cancellation,
  so unlike the other fix this directly costs ~20s more wall-clock time,
  but that's inherent to what it's testing.

internal/agenteval/benchmark_test.go

@@ -455,11 +455,15 @@ func TestHarnessTimeoutCancelsBlockedAgent(t *testing.T) {
 
 	// The timeout is longer than materialization of the small fixture even on
 	// slower Windows CI runners, so the agent is reached before it blocks until
-	// the deadline fires.
+	// the deadline fires. 10s wasn't generous enough in practice — it failed
+	// on an otherwise-unrelated PR merge (materialization alone consumed the
+	// full budget on a contended Windows runner, so the agent was never
+	// reached); widened for headroom, since the assertions below still bound
+	// worst-case test time via that same deadline, not this constant.
 	report := harness.Run(context.Background(), suitePath, suite, BenchmarkInput{
 		TaskID:   "document-stream-json-verify-events",
 		WorkRoot: t.TempDir(),
-		Timeout:  10 * time.Second,
+		Timeout:  30 * time.Second,
 	})
 
 	if !agentReached {

internal/specialist/output_tool_test.go

@@ -124,10 +124,16 @@ func TestOutputToolBlocksUntilTaskCompletes(t *testing.T) {
 	tool := NewOutputTool(manager)
 	tool.PollInterval = time.Millisecond
 
+	// timeout is a ceiling, not a target: the test still returns as soon as the
+	// 5ms write lands (PollInterval keeps it responsive), it just tolerates a
+	// slow/contended CI runner without failing. A 100ms budget against a 5ms
+	// write only left 20x margin, which Windows CI runners (coarser default
+	// timer granularity, heavier scheduling contention than Linux/macOS) blew
+	// through and failed the test on an otherwise-unrelated PR merge.
 	result := tool.Run(context.Background(), map[string]any{
 		"task_id": "child_task",
 		"block":   true,
-		"timeout": 100,
+		"timeout": 2000,
 	})
 
 	if result.Status != tools.StatusOK || !strings.Contains(result.Output, "output:\nfinished") {