diff --git a/idf_zacus/components/game_endpoint/CMakeLists.txt b/idf_zacus/components/game_endpoint/CMakeLists.txt
index e3b3f0c..dd44449 100644
--- a/idf_zacus/components/game_endpoint/CMakeLists.txt
+++ b/idf_zacus/components/game_endpoint/CMakeLists.txt
@@ -15,6 +15,7 @@ idf_component_register(
         log
         joltwallet__littlefs
         puzzle_state
+        media_manager
     PRIV_REQUIRES
         local_puzzles
 )
diff --git a/idf_zacus/components/game_endpoint/game_endpoint.c b/idf_zacus/components/game_endpoint/game_endpoint.c
index 26b7d31..a3657bb 100644
--- a/idf_zacus/components/game_endpoint/game_endpoint.c
+++ b/idf_zacus/components/game_endpoint/game_endpoint.c
@@ -18,6 +18,7 @@
 #include "esp_http_server.h"
 #include "esp_littlefs.h"
 #include "esp_log.h"
+#include "media_manager.h"
 #include "esp_system.h"
 #include "freertos/FreeRTOS.h"
 #include "freertos/task.h"
@@ -866,6 +867,55 @@ static esp_err_t handle_step_post(httpd_req_t *req) {
     }
 }
 
+// ─── POST /game/media/play — trigger a media_manager WAV cue ─────────────────
+// Body: { "path": "apps/<id>/cue.wav" | "/littlefs/music/foo.wav" }. Streams
+// the file to the MAX98357A via media_manager (P4 real playback). Lets the
+// gateway / atelier fire audio cues without an NPC decision.
+static esp_err_t handle_media_play_post(httpd_req_t *req) {
+    char body[513] = {0};
+    if (req->content_len <= 0 || req->content_len > 512) {
+        return send_error(req, "400 Bad Request", "body must be 1..512 bytes");
+    }
+    int total = 0;
+    while (total < (int) req->content_len) {
+        int got = httpd_req_recv(req, body + total, req->content_len - total);
+        if (got <= 0) {
+            if (got == HTTPD_SOCK_ERR_TIMEOUT) continue;
+            return send_error(req, "400 Bad Request", "recv failed");
+        }
+        total += got;
+    }
+    body[total] = '\0';
+
+    cJSON *root = cJSON_Parse(body);
+    if (!root) {
+        return send_error(req, "400 Bad Request", "malformed json");
+    }
+    const cJSON *path_j = cJSON_GetObjectItemCaseSensitive(root, "path");
+    if (!cJSON_IsString(path_j) || !path_j->valuestring || path_j->valuestring[0] == '\0') {
+        cJSON_Delete(root);
+        return send_error(req, "400 Bad Request", "missing or empty path");
+    }
+    char path[160];
+    strncpy(path, path_j->valuestring, sizeof(path) - 1);
+    path[sizeof(path) - 1] = '\0';
+    cJSON_Delete(root);
+
+    esp_err_t err = media_manager_play(path);
+    if (err == ESP_OK) {
+        char buf[200];
+        snprintf(buf, sizeof(buf), "{\"status\":\"playing\",\"path\":\"%s\"}", path);
+        return send_json(req, "200 OK", buf);
+    }
+    if (err == ESP_ERR_NOT_FOUND) {
+        return send_error(req, "404 Not Found", "file_not_found");
+    }
+    if (err == ESP_ERR_INVALID_ARG) {
+        return send_error(req, "400 Bad Request", "invalid_path");
+    }
+    return send_error(req, "500 Internal Server Error", esp_err_to_name(err));
+}
+
 // ─── GET /game/puzzle_state ──────────────────────────────────────────────────
 //
 // Returns {"step_id":"STEP_X"|null, "solved":[1,3], "code":"125"}.
@@ -1084,6 +1134,12 @@ esp_err_t game_endpoint_init(httpd_handle_t server) {
         .handler  = handle_espnow_cmd_post,
         .user_ctx = NULL,
     };
+    static const httpd_uri_t uri_media_play_post = {
+        .uri      = "/game/media/play",
+        .method   = HTTP_POST,
+        .handler  = handle_media_play_post,
+        .user_ctx = NULL,
+    };
 
     // Bring up the ESP-NOW mesh transport. The master is primarily a sender
     // (the relay handler) but we also register the apply adapter so a peer
@@ -1156,6 +1212,10 @@ esp_err_t game_endpoint_init(httpd_handle_t server) {
     if (err != ESP_OK) {
         ESP_LOGW(TAG, "register GET /game/puzzle_state: %s", esp_err_to_name(err));
     }
+    err = httpd_register_uri_handler(server, &uri_media_play_post);
+    if (err != ESP_OK) {
+        ESP_LOGW(TAG, "register POST /game/media/play: %s", esp_err_to_name(err));
+    }
 
     ESP_LOGI(TAG, "game endpoint registered "
                   "(GET+POST /game/group_profile, POST /game/scenario%s, "
diff --git a/idf_zacus/components/media_manager/CMakeLists.txt b/idf_zacus/components/media_manager/CMakeLists.txt
index fa9c79a..bd3a38a 100644
--- a/idf_zacus/components/media_manager/CMakeLists.txt
+++ b/idf_zacus/components/media_manager/CMakeLists.txt
@@ -17,5 +17,7 @@ idf_component_register(
         driver
         esp_timer
         esp_system
+        freertos
+        voice_pipeline
         joltwallet__littlefs
 )
diff --git a/idf_zacus/components/media_manager/media_manager.c b/idf_zacus/components/media_manager/media_manager.c
index 206c4ec..057b56b 100644
--- a/idf_zacus/components/media_manager/media_manager.c
+++ b/idf_zacus/components/media_manager/media_manager.c
@@ -31,11 +31,18 @@
 #include <sys/types.h>
 #include <unistd.h>
 
+#include "freertos/FreeRTOS.h"
+#include "freertos/task.h"
+
 #include "esp_check.h"
 #include "esp_err.h"
 #include "esp_log.h"
 #include "esp_timer.h"
 
+// Slice P4: real WAV playback streams 16-bit PCM to the MAX98357A over the
+// I2S TX channel already owned by voice_pipeline (shared DAC, no codec init).
+#include "voice_pipeline.h"
+
 static const char *TAG = "media_manager";
 
 #define MEDIA_DEFAULT_MUSIC_DIR    "/littlefs/music"
@@ -59,6 +66,8 @@ static struct {
     uint32_t                   playback_ends_ms;
     FILE                      *recording_file;
     uint32_t                   recording_data_bytes;
+    TaskHandle_t               playback_task;   // P4: real WAV streamer, NULL when idle
+    volatile bool              playback_stop;   // request the streamer to abort
 } s_media;
 
 // ─── Helpers ─────────────────────────────────────────────────────────────────
@@ -332,6 +341,135 @@ void media_manager_note_step_change(void) {
     }
 }
 
+// ─── P4: real WAV playback over the shared MAX98357A I2S TX ──────────────────
+//
+// Streams 16-bit PCM WAV from LittleFS to the DAC via voice_pipeline's TX
+// channel. WAV (not MP3) keeps this decoder-free; MP3 stays a later slice.
+// NOTE: shares the I2S TX with TTS — callers must not overlap a `speak_*`
+// exchange with music playback (the NPC coordination layer already serialises
+// cues). A TX mutex is the natural follow-up.
+
+#define MEDIA_PLAY_CHUNK_BYTES   1024U
+
+typedef struct {
+    uint32_t sample_rate;
+    uint16_t channels;
+    uint16_t bits;
+    uint32_t data_size;   // PCM bytes in the data chunk
+} wav_info_t;
+
+static uint32_t rd_u32(const uint8_t *p) {
+    return (uint32_t) p[0] | ((uint32_t) p[1] << 8) |
+           ((uint32_t) p[2] << 16) | ((uint32_t) p[3] << 24);
+}
+
+// Parse a RIFF/WAVE header, leaving `f` positioned at the PCM data. Accepts
+// 16-bit PCM mono/stereo only; returns false otherwise.
+static bool parse_wav_header(FILE *f, wav_info_t *out) {
+    uint8_t riff[12];
+    if (fread(riff, 1, 12, f) != 12) return false;
+    if (memcmp(riff, "RIFF", 4) != 0 || memcmp(riff + 8, "WAVE", 4) != 0) return false;
+
+    bool have_fmt = false;
+    uint8_t ck[8];
+    while (fread(ck, 1, 8, f) == 8) {
+        uint32_t csz = rd_u32(ck + 4);
+        if (memcmp(ck, "fmt ", 4) == 0) {
+            uint8_t fmt[16];
+            uint32_t want = csz < sizeof(fmt) ? csz : (uint32_t) sizeof(fmt);
+            if (fread(fmt, 1, want, f) != want) return false;
+            if (((uint16_t) fmt[0] | ((uint16_t) fmt[1] << 8)) != 1U) return false; // PCM
+            out->channels    = (uint16_t) fmt[2] | ((uint16_t) fmt[3] << 8);
+            out->sample_rate = rd_u32(fmt + 4);
+            out->bits        = (uint16_t) fmt[14] | ((uint16_t) fmt[15] << 8);
+            if (csz > want) fseek(f, (long) (csz - want), SEEK_CUR);
+            have_fmt = true;
+        } else if (memcmp(ck, "data", 4) == 0) {
+            out->data_size = csz;
+            return have_fmt && out->bits == 16U &&
+                   (out->channels == 1U || out->channels == 2U);
+        } else {
+            fseek(f, (long) (csz + (csz & 1U)), SEEK_CUR); // skip + word-align pad
+        }
+    }
+    return false;
+}
+
+static void apply_volume(int16_t *s, size_t count, uint8_t volume) {
+    if (volume >= 100U) return;
+    for (size_t i = 0; i < count; ++i) {
+        s[i] = (int16_t) (((int32_t) s[i] * (int32_t) volume) / 100);
+    }
+}
+
+static void clear_playback_state(void) {
+    s_media.snapshot.playing = false;
+    s_media.snapshot.playing_path[0] = '\0';
+    s_media.playback_ends_ms = 0U;
+    s_media.playback_task = NULL;
+}
+
+static void playback_task(void *arg) {
+    (void) arg;
+    char path[MEDIA_PATH_MAX];
+    copy_text(path, sizeof(path), s_media.snapshot.playing_path);
+
+    FILE *f = fopen(path, "rb");
+    wav_info_t wav = {0};
+    if (f == NULL || !parse_wav_header(f, &wav)) {
+        if (f) fclose(f);
+        ESP_LOGE(TAG, "playback: %s is not a 16-bit PCM WAV", path);
+        set_last_error("media_play_bad_wav");
+        clear_playback_state();
+        vTaskDelete(NULL);
+        return;
+    }
+
+    ESP_LOGI(TAG, "playback start: %s (%lu Hz, %uch, %lu B) vol=%u",
+             path, (unsigned long) wav.sample_rate, wav.channels,
+             (unsigned long) wav.data_size, s_media.volume);
+
+    esp_err_t err = voice_pipeline_play_start(wav.sample_rate, "wav");
+    if (err != ESP_OK) {
+        ESP_LOGE(TAG, "play_start failed: %s", esp_err_to_name(err));
+        fclose(f);
+        set_last_error("media_play_i2s");
+        clear_playback_state();
+        vTaskDelete(NULL);
+        return;
+    }
+
+    uint8_t buf[MEDIA_PLAY_CHUNK_BYTES];
+    uint32_t remaining = wav.data_size;
+    uint32_t written = 0;
+    while (remaining > 0U && !s_media.playback_stop) {
+        size_t want = remaining < sizeof(buf) ? remaining : sizeof(buf);
+        size_t n = fread(buf, 1, want, f);
+        if (n == 0U) break;
+        remaining -= (uint32_t) n;
+
+        size_t pcm = n;
+        if (wav.channels == 2U) {              // downmix interleaved stereo16 → mono16
+            int16_t *s = (int16_t *) buf;
+            size_t frames = n / 4U;
+            for (size_t i = 0; i < frames; ++i) {
+                s[i] = (int16_t) (((int32_t) s[2 * i] + s[2 * i + 1]) / 2);
+            }
+            pcm = frames * 2U;
+        }
+        apply_volume((int16_t *) buf, pcm / 2U, s_media.volume);
+        if (voice_pipeline_play_chunk(buf, pcm) != ESP_OK) break;
+        written += (uint32_t) pcm;
+    }
+
+    voice_pipeline_play_end();
+    fclose(f);
+    ESP_LOGI(TAG, "playback done: %s (%lu B to DAC%s)", path,
+             (unsigned long) written, s_media.playback_stop ? ", stopped" : "");
+    clear_playback_state();
+    vTaskDelete(NULL);
+}
+
 esp_err_t media_manager_play(const char *path) {
     if (!s_media.initialized) {
         return ESP_ERR_INVALID_STATE;
@@ -354,15 +492,27 @@ esp_err_t media_manager_play(const char *path) {
         return ESP_ERR_NOT_FOUND;
     }
 
-    // TODO(slice-4+): hand off to the real I2S MP3 decoder here.
-    const uint32_t now_ms = (uint32_t) (esp_timer_get_time() / 1000LL);
+    // Stop any in-flight playback before starting the next cue.
+    if (s_media.playback_task != NULL) {
+        (void) media_manager_stop();
+    }
+
+    s_media.playback_stop = false;
     s_media.snapshot.playing = true;
     copy_text(s_media.snapshot.playing_path,
               sizeof(s_media.snapshot.playing_path), resolved);
-    s_media.playback_ends_ms = now_ms + MEDIA_STUB_PLAYBACK_MS;
+    s_media.playback_ends_ms = 0U;   // real playback: tick must not auto-clear
     clear_last_error();
-    ESP_LOGI(TAG, "playing %s (simulated %ums) vol=%u",
-             resolved, MEDIA_STUB_PLAYBACK_MS, s_media.volume);
+
+    BaseType_t ok = xTaskCreate(playback_task, "media_play", 5120, NULL, 4,
+                                &s_media.playback_task);
+    if (ok != pdPASS) {
+        s_media.playback_task = NULL;
+        s_media.snapshot.playing = false;
+        s_media.snapshot.playing_path[0] = '\0';
+        set_last_error("media_play_task");
+        return ESP_ERR_NO_MEM;
+    }
     return ESP_OK;
 }
 
@@ -370,6 +520,12 @@ esp_err_t media_manager_stop(void) {
     if (!s_media.initialized) {
         return ESP_ERR_INVALID_STATE;
     }
+    if (s_media.playback_task != NULL) {
+        s_media.playback_stop = true;
+        for (int i = 0; i < 100 && s_media.playback_task != NULL; ++i) {
+            vTaskDelay(pdMS_TO_TICKS(10));   // let the streamer drain + self-delete
+        }
+    }
     if (s_media.snapshot.playing) {
         ESP_LOGI(TAG, "stop %s", s_media.snapshot.playing_path);
     }