fix(multimodal): stabilize downstream audio playback

- add configurable prebuffer and adaptive rebuffer for packet jitter
- mute stalled output, auto-clear I2S DMA, and abort inactive playback after 3 seconds
- discard stale-generation packets and add optional downstream diagnostics

main/Kconfig.projbuild

@@ -1101,6 +1101,54 @@ menu "Example Configuration"
                     the ES8311 output path. This keeps the PA control behavior
                     unchanged while suppressing idle speaker noise.
 
+            config MULTIMODAL_AUDIO_DOWNSTREAM_ACTIVE_STALL_MUTE_MS
+                int "Downstream active response stall timeout (ms)"
+                default 3000
+                range 500 10000
+                help
+                    Abort the current playback generation when neither audio
+                    output nor downstream response packets advance for this
+                    long. The output is muted first so a stalled response
+                    cannot leave stale DMA audio audible indefinitely.
+
+            config MULTIMODAL_AUDIO_DOWNSTREAM_PREBUFFER_MS
+                int "Downstream playback prebuffer (ms)"
+                default 160
+                range 0 1000
+                help
+                    Audio accumulated before initial playback. This adds a
+                    small first-audio delay in exchange for absorbing normal
+                    Wi-Fi and server delivery jitter.
+
+            config MULTIMODAL_AUDIO_DOWNSTREAM_REBUFFER_GAP_MS
+                int "Downstream underrun detection delay (ms)"
+                default 120
+                range 20 500
+                help
+                    Continuous empty-queue interval that mutes the output and
+                    returns playback to buffering instead of restarting on
+                    every isolated packet.
+
+            config MULTIMODAL_AUDIO_DOWNSTREAM_MAX_REBUFFER_MS
+                int "Maximum adaptive downstream recovery buffer (ms)"
+                default 640
+                range 160 2000
+                help
+                    Maximum audio accumulated after repeated queue underruns.
+                    The recovery target grows from twice the initial prebuffer
+                    up to this limit for an unstable response.
+
+            config MULTIMODAL_AUDIO_DOWNSTREAM_DIAGNOSTICS_ENABLE
+                bool "Enable downstream audio diagnostics"
+                default n
+                help
+                    Collect and report low-rate downstream audio timing, queue
+                    depth, queue starvation, enqueue failure, and playback
+                    failure statistics. Active responses also report a
+                    once-per-second wait snapshot when no audio advances.
+                    Disabled by default to avoid diagnostic logging and
+                    per-packet accounting in production builds.
+
             config MULTIMODAL_AUDIO_DOWNSTREAM_PLAYBACK_TASK_STACK
                 int "Downstream playback task stack (bytes)"
                 default 12288