fix(audio): hardware-paced audio (no wall-clock silence mixing) + aresample=async to lock A/V sync

services/capture/deltacast-bridge/main.c

@@ -252,15 +252,32 @@ static void *audio_thread(void *arg) {
         #endif
         fcntl(fd, F_SETPIPE_SZ, 1024 * 1024);
 
-        /* Reset wall-clock baseline after potentially blocking on open(). */
+        /* Reset wall-clock baseline after potentially blocking on open().
+         * Only used for the SILENCE fallback path (no hardware audio). */
         struct timespec next;
         clock_gettime(CLOCK_MONOTONIC, &next);
 
+        /* Audio-rate telemetry: count samples written per second so drift is
+         * visible in the log ([audio:N] rate=<samples/s>). At 48 kHz it must
+         * average 48000. */
+        unsigned long dbg_samples = 0;
+        struct timespec dbg_t0; clock_gettime(CLOCK_MONOTONIC, &dbg_t0);
+
         /* Inner loop: feed audio into the open FIFO until reader exits (EPIPE). */
         while (!atomic_load(&g_stop) && !atomic_load(&g_port_stop[ps->port])) {
             size_t out_bytes = 0;
 
             if (have_vhd_audio) {
+                /* HARDWARE-PACED PATH (the normal case).
+                 * VHD_LockSlotHandle blocks until the board has the next audio
+                 * slot ready — this slot is generated from the SAME SDI signal
+                 * as the video, so blocking here paces audio in lockstep with
+                 * video at the TRUE hardware rate. We write ONLY the real
+                 * samples the board gives us (no silence padding, no wall-clock
+                 * sleep) so the audio timeline length exactly tracks video.
+                 * This is the fix for progressive A/V drift: mixing wall-clock
+                 * paced silence with variable-length real reads made the audio
+                 * stream length diverge from the video stream length. */
                 r = VHD_LockSlotHandle(stream, &slot);
                 if (r == VHDERR_NOERROR) {
                     ai.pAudioGroups[0].pAudioChannels[0].DataSize = (ULONG)buf_sz;
@@ -269,26 +286,46 @@ static void *audio_thread(void *arg) {
                         if (sz > 0 && (size_t)sz <= buf_sz) out_bytes = (size_t)sz;
                     }
                     VHD_UnlockSlotHandle(slot);
-                } else if (r != VHDERR_TIMEOUT) {
+
+                    if (out_bytes > 0) {
+                        if (write_all(fd, buf, out_bytes) < 0) {
+                            fprintf(stderr, "[audio:%u] EPIPE — waiting for next reader\n", ps->port);
+                            break;
+                        }
+                        dbg_samples += out_bytes / FRAME_BYTES;
+                        struct timespec dnow; clock_gettime(CLOCK_MONOTONIC, &dnow);
+                        double el = (dnow.tv_sec - dbg_t0.tv_sec) + (dnow.tv_nsec - dbg_t0.tv_nsec)/1e9;
+                        if (el >= 5.0) {
+                            fprintf(stderr, "[audio:%u] rate=%.1f samples/s (target 48000)\n",
+                                    ps->port, dbg_samples / el);
+                            dbg_samples = 0; dbg_t0 = dnow;
+                        }
+                    }
+                    /* No wall-clock sleep — the board's slot cadence is the clock. */
+                    continue;
+                } else if (r == VHDERR_TIMEOUT) {
+                    /* No slot yet — loop and try again (do NOT inject silence,
+                     * that would add extra samples and cause drift). */
+                    continue;
+                } else {
                     fprintf(stderr, "[audio:%u] lock error %lu — degrading to silence\n",
                             ps->port, r);
                     VHD_StopStream(stream);
                     VHD_CloseStreamHandle(stream);
                     stream = NULL;
                     have_vhd_audio = 0;
+                    clock_gettime(CLOCK_MONOTONIC, &next); /* rebase silence clock */
                 }
             }
 
-            if (out_bytes == 0) {
-                memset(buf, 0, tick_bytes);
-                out_bytes = tick_bytes;
-            }
+            /* SILENCE FALLBACK PATH (no hardware audio available).
+             * Wall-clock paced one-frame-of-silence per video-frame interval so
+             * ffmpeg's input 1 never starves and audio length still tracks
+             * real time. */
+            memset(buf, 0, tick_bytes);
+            out_bytes = tick_bytes;
 
             if (write_all(fd, buf, out_bytes) < 0) {
-                /* EPIPE: ffmpeg reader on this port died (session stop/restart).
-                 * Close and break to the outer loop which will reopen and block
-                 * until the next ffmpeg reader connects.
-                 * Do NOT set g_stop — other ports must keep running. */
                 fprintf(stderr, "[audio:%u] EPIPE — waiting for next reader\n", ps->port);
                 break;
             }

services/capture/src/capture-manager.js

@@ -1008,6 +1008,11 @@ exit "$BMXRC"
           '-filter_complex', filterStr,
           // Output 0 — ProRes/MOV master (local temp, uploaded to S3 on stop)
           '-map', '[vhi]', '-map', audioMap,
+          // Keep raw audio aligned to the video clock. The two raw FIFOs carry
+          // no timestamps; -af aresample=async lets ffmpeg stretch/squeeze audio
+          // to correct any tiny rate mismatch so A/V never drifts over a long
+          // take. Applies to this output's mapped audio stream.
+          '-af', 'aresample=async=1:min_hard_comp=0.100000:first_pts=0',
           ...hiresCodecArgs,
           hiresOutput,
           // Output 1 — low-latency H.264 HLS preview for the UI monitor.