switched rover <-> server streams to rtsptcp, and moved mediamtx to be a server owned and configured child process!

pi/bin/audio-forward-listener.sh

@@ -9,9 +9,8 @@ if [[ ! -f "$ENV_FILE" ]]; then
   exit 1
 fi
 
-# Load KEY=VALUE pairs from media.env without evaluating shell syntax. The
-# forward URL is data produced by roverd, and treating it as shell code would
-# break on normal SRT query-string characters such as '&'.
+# Load KEY=VALUE pairs from media.env without evaluating shell syntax. The forward URL is
+# data produced by roverd and must never be interpreted as executable shell code.
 load_env_file() {
   local content=""
 
@@ -95,6 +94,9 @@ run_pipeline() {
     -flags low_delay
     -analyzeduration 200k
     -probesize 32k
+    # The forwarded-audio URL is RTSP. Pinning TCP avoids ffmpeg negotiating the
+    # separate unreliable RTP/UDP transport that the server intentionally disables.
+    -rtsp_transport tcp
     -i "${ROVERD_AUDIO_PLAYBACK_FORWARD_URL}"
     -vn
   )

pi/bin/audio-only-publisher.sh

@@ -9,9 +9,8 @@ if [[ ! -f "$ENV_FILE" ]]; then
 	exit 1
 fi
 
-# Load KEY=VALUE pairs from media.env without evaluating shell syntax. SRT URLs
-# contain characters such as '&' and '#!', so sourcing this file would treat a
-# data file as code and can split a valid URL into shell control operators.
+# Load KEY=VALUE pairs from media.env without evaluating shell syntax. URLs are data;
+# sourcing this file would unnecessarily treat server-provided values as shell code.
 load_env_file() {
 	local content=""
 
@@ -123,13 +122,14 @@ run_pipeline() {
 		-frame_duration 20
 		-compression_level 0
 
-		# Mirror the video publisher's MPEG-TS low-latency settings. Without
-		# these, ffmpeg is allowed to hold packets for mux timing, which is
-		# exactly the wrong tradeoff for live rover feedback.
+		# RTSP carries the existing Opus stream directly, avoiding MediaMTX's costly
+		# MPEG-TS demux without changing microphone capture or encoding quality. TCP is
+		# required for the same reliable local-network behavior as the video publisher.
 		-flush_packets 1
 		-muxdelay 0
 		-muxpreload 0
-		-f mpegts
+		-f rtsp
+		-rtsp_transport tcp
 		"${ROVERD_AUDIO_CAPTURE_PUBLISH_URL}"
 	)
 

pi/bin/debian-laptop-video-publisher.sh

@@ -9,9 +9,8 @@ if [[ ! -f "$ENV_FILE" ]]; then
 	exit 1
 fi
 
-# Load roverd's generated media.env as data instead of sourcing it as shell.
-# The SRT publish URL contains normal query-string characters like '&' and '#!',
-# so evaluating the file would be both fragile and unnecessary.
+# Load roverd's generated media.env as data instead of sourcing it as shell. URLs are
+# configuration data, so evaluating the file would be both fragile and unnecessary.
 load_env_file() {
 	local content=""
 
@@ -103,6 +102,8 @@ if [[ "${ROVERD_VIDEO_INVERT}" -ne 0 ]]; then
 fi
 
 run_pipeline() {
+	# Keep laptop rovers on the same transport contract as Pi camera rovers. This changes
+	# only the encoded stream's carrier; V4L2 capture and H264 encoding remain untouched.
 	"${FFMPEG_BIN_PATH}" \
 		-hide_banner \
 		-loglevel warning \
@@ -130,7 +131,8 @@ run_pipeline() {
 		-flush_packets 1 \
 		-muxdelay 0 \
 		-muxpreload 0 \
-		-f mpegts \
+		-f rtsp \
+		-rtsp_transport tcp \
 		"${ROVERD_VIDEO_PUBLISH_URL}"
 }
 

pi/bin/media-transport.test.sh

@@ -0,0 +1,37 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# These publishers contain hardware-facing infinite retry loops, so executing them in a unit
+# test would require unsafe process-group traps and fake camera/ALSA devices. Pin the small
+# transport boundary directly instead: every publisher must request RTSP/TCP and none may
+# reintroduce the high-latency MPEG-TS muxer.
+SCRIPT_DIR=$(cd "$(dirname "$0")" && pwd)
+
+assert_rtsp_tcp() {
+	local file="$1"
+	if ! grep -q -- '-f rtsp' "$file"; then
+		echo "Missing RTSP muxer in $file" >&2
+		exit 1
+	fi
+	if ! grep -q -- '-rtsp_transport tcp' "$file"; then
+		echo "Missing RTSP/TCP pin in $file" >&2
+		exit 1
+	fi
+	if grep -q -- '-f mpegts' "$file"; then
+		echo "Unexpected MPEG-TS muxer in $file" >&2
+		exit 1
+	fi
+}
+
+assert_rtsp_tcp "$SCRIPT_DIR/video-publisher.sh"
+assert_rtsp_tcp "$SCRIPT_DIR/debian-laptop-video-publisher.sh"
+assert_rtsp_tcp "$SCRIPT_DIR/audio-only-publisher.sh"
+
+# The speaker path reads rather than publishes, so it has no output muxer. It must still pin
+# RTSP/TCP before its input URL to match the server's TCP-only listener.
+if ! grep -q -- '-rtsp_transport tcp' "$SCRIPT_DIR/audio-forward-listener.sh"; then
+	echo "Missing RTSP/TCP input pin in audio-forward-listener.sh" >&2
+	exit 1
+fi
+
+echo "Media publisher transport checks passed"

pi/bin/video-publisher.sh

@@ -110,6 +110,10 @@ else
 fi
 
 run_pipeline() {
+  # MPEG-TS added most of the former rover-to-browser latency inside MediaMTX's
+  # demuxer. RTSP carries the same encoded H264 without changing the camera or codec.
+  # TCP is explicit because plain RTSP/RTP over UDP has no retransmission and proved
+  # unreliable even though MediaMTX still reported the incomplete stream as ready.
   "${LIBCAMERA_BIN_PATH}" \
     --inline \
     --timeout 0 \
@@ -142,7 +146,8 @@ run_pipeline() {
       -flush_packets 1 \
       -muxdelay 0 \
       -muxpreload 0 \
-      -f mpegts \
+      -f rtsp \
+      -rtsp_transport tcp \
       "${ROVERD_VIDEO_PUBLISH_URL}"
 }