diff --git a/docs/superpowers/plans/2026-08-28-ci-main-green.md b/docs/superpowers/plans/2026-08-28-ci-main-green.md
index ce851b77..cc284c80 100644
--- a/docs/superpowers/plans/2026-08-28-ci-main-green.md
+++ b/docs/superpowers/plans/2026-08-28-ci-main-green.md
@@ -10,6 +10,13 @@
 
 **Spec:** `docs/TECHNICAL-RISKS.md` and `AGENTS.md`.
 
+The shared HTTP muxer codec-epoch regression must establish the direct AAC
+output before injecting the next source codec. FLV and TS emit that output
+immediately, while fMP4 requires the DTS=500 video boundary to flush its
+fragment; the test therefore waits after that boundary and before G.711
+input. This keeps the test independent of goroutine scheduling while still
+covering the direct-to-transformed transition.
+
 ## Global Constraints
 
 - Use Go 1.26 or newer and the repository's `audiocodec` build tag for the tagged baseline.
diff --git a/module/httpstream/muxer_worker_audiocodec_test.go b/module/httpstream/muxer_worker_audiocodec_test.go
index 6c5fb511..ce95e726 100644
--- a/module/httpstream/muxer_worker_audiocodec_test.go
+++ b/module/httpstream/muxer_worker_audiocodec_test.go
@@ -272,6 +272,11 @@ func TestSharedMuxersSwitchFromDirectAACToTransformedAudio(t *testing.T) {
 				avframe.MediaTypeVideo, avframe.CodecH264, avframe.FrameTypeInterframe,
 				500, 500, []byte{0, 0, 0, 2, 0x41, 0x05},
 			))
+			// Establish the direct-AAC epoch before introducing an incompatible
+			// source codec. The live and transformed readers share an unbuffered
+			// delivery channel, so writing both epochs back-to-back would let
+			// scheduler timing discard an initial direct frame.
+			packets = waitForMuxerPayloads(t, format, initData, reader, packets, directAAC)
 			writeG711MuxerFrames(stream, 1000, 20, 0xff)
 			stream.WriteFrame(avframe.NewAVFrame(
 				avframe.MediaTypeVideo, avframe.CodecH264, avframe.FrameTypeInterframe,