diff --git a/docs/plans/v1/_files/internal/limiter/limiter_test.go b/docs/plans/v1/_files/internal/limiter/limiter_test.go
index bf90588..6ab30e2 100644
--- a/docs/plans/v1/_files/internal/limiter/limiter_test.go
+++ b/docs/plans/v1/_files/internal/limiter/limiter_test.go
@@ -32,10 +32,10 @@ func TestParallelAndQueue(t *testing.T) {
 	go func() {
 		rel, waited, err := l.Acquire(ctx, "alpha", "m")
 		if err == nil {
-			defer rel()
 			if waited < 40*time.Millisecond {
 				err = errors.New("third acquire did not wait")
 			}
+			rel() // release before reporting, so the final count check cannot race it
 		}
 		got3 <- err
 	}()
diff --git a/docs/plans/v1/_files/internal/proxy/proxy_test.go b/docs/plans/v1/_files/internal/proxy/proxy_test.go
index c1bb26a..39980aa 100644
--- a/docs/plans/v1/_files/internal/proxy/proxy_test.go
+++ b/docs/plans/v1/_files/internal/proxy/proxy_test.go
@@ -85,6 +85,19 @@ func TestDifferentConversationsSpreadByFreeSlots(t *testing.T) {
 	}
 }
 
+// waitUntil polls cond every 5 ms for up to two seconds and fails the test if it never holds.
+func waitUntil(t *testing.T, cond func() bool) {
+	t.Helper()
+	deadline := time.Now().Add(2 * time.Second)
+	for time.Now().Before(deadline) {
+		if cond() {
+			return
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+	t.Fatal("condition not reached within two seconds")
+}
+
 func TestQueueFullIs503(t *testing.T) {
 	alpha := newUpstream(t, "alpha")
 	alpha.delay = 400 * time.Millisecond
@@ -99,14 +112,20 @@ hosts = ["alpha"]
 default_model = "shared"
 `, alpha)
 	codes := make(chan int, 3)
-	for i := 1; i <= 3; i++ {
-		go func(i int) {
+	fire := func(i int) {
+		go func() {
 			resp := r.post("/r/v1/chat/completions", conversation(i, 1))
 			drain(resp)
 			codes <- resp.StatusCode
-		}(i)
-		time.Sleep(30 * time.Millisecond) // arrival order: 1 runs, 2 queues, 3 finds the queue full
+		}()
 	}
+	// Arrival order is enforced by watching the limiter, not by sleeping: 1 runs, 2 queues,
+	// 3 finds the queue full.
+	fire(1)
+	waitUntil(t, func() bool { return r.lim.InFlight("alpha", "shared") == 1 })
+	fire(2)
+	waitUntil(t, func() bool { return r.lim.Queued("alpha", "shared") == 1 })
+	fire(3)
 	got := map[int]int{}
 	for i := 0; i < 3; i++ {
 		got[<-codes]++
diff --git a/docs/plans/v2/README.md b/docs/plans/v2/README.md
index 5194a88..8e5c75a 100644
--- a/docs/plans/v2/README.md
+++ b/docs/plans/v2/README.md
@@ -82,3 +82,17 @@ reason.
   per argv element, so check 3 could never pass. Test fault (mine): the body now goes through a
   file (`-d @file`). Ornith diagnosed it correctly. Resumed from the working tree with the
   corrected script.
+- 2026-09-25, task 05, third session: `TestParallelAndQueue` (v1 given `limiter_test.go`) failed
+  once under full-suite `-race` load with `after releases: inflight 1 queued 0`. Test fault
+  (mine): the third acquirer sent its result before its deferred release ran, so the final
+  count check could observe one slot still held. The given file now releases before reporting.
+  Ornith found it and measured the flake rate rather than editing the protected file.
+- 2026-09-25, task 05, third session: also saw `TestQueueFullIs503` fail with
+  `Requests:3 Errors:2`. Two causes. (1) Test fault (mine): arrival order rested on 30 ms
+  sleeps; the given test now waits on the limiter's in-flight and queued counts. (2) A real v1
+  defect, verified by the owner with a diagnostic build (5 of 8 runs): after a forward completes,
+  `forward.go` checks `r.Context().Err()` and, when the client has already closed its connection,
+  records a served 200 as a 499 "client cancelled" error. Cancellation must be what the reverse
+  proxy itself observed, never a post-hoc context check. Scheduled as v2.1 task 01; not fixed in
+  task 05, which is wiring only. The session then ended on a refused read of `/proc/loadavg` —
+  the seventh refusal-ending. Model fault.