[fix](trx-server): chunk audio history replay to stay under 16 MiB cap

The compressed history blob sent on each audio client connect could exceed the client's MAX_HISTORY_PAYLOAD_SIZE (16 MiB) once enough decoder records accumulated, causing the client to reject the frame, drop the connection, and reconnect — producing a 1 Hz reconnect storm.

Walk the framed blob at message boundaries and emit one AUDIO_MSG_HISTORY_COMPRESSED per ~8 MiB uncompressed chunk. The split is wire-compatible: clients already loop on read_audio_msg and process each history frame independently.

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>

src/trx-server/src/audio.rs

@@ -200,6 +200,41 @@ impl StreamErrorLogger {
     }
 }
 
+/// Walk a length-prefixed history blob and return the byte ranges of each
+/// chunk to send. Each record is `[u8 type][u32 BE len][payload]`. A chunk is
+/// flushed when adding the next record would push it past `threshold`. A
+/// single record larger than `threshold` becomes its own chunk — the wire-side
+/// `MAX_HISTORY_PAYLOAD_SIZE` cap is generous enough for any realistic decode.
+/// A malformed tail (truncated header or payload) is dropped silently; the
+/// blob is self-built so this is defence in depth, not error reporting.
+fn split_history_chunks(blob: &[u8], threshold: usize) -> Vec<std::ops::Range<usize>> {
+    let mut chunks = Vec::new();
+    let mut chunk_start = 0usize;
+    let mut pos = 0usize;
+    while pos < blob.len() {
+        let len_end = pos.saturating_add(5);
+        if len_end > blob.len() {
+            break;
+        }
+        let payload_len =
+            u32::from_be_bytes([blob[pos + 1], blob[pos + 2], blob[pos + 3], blob[pos + 4]])
+                as usize;
+        let msg_end = len_end.saturating_add(payload_len);
+        if msg_end > blob.len() {
+            break;
+        }
+        if pos > chunk_start && (msg_end - chunk_start) > threshold {
+            chunks.push(chunk_start..pos);
+            chunk_start = pos;
+        }
+        pos = msg_end;
+    }
+    if chunk_start < pos {
+        chunks.push(chunk_start..pos);
+    }
+    chunks
+}
+
 fn classify_stream_error(err: &str) -> &'static str {
     if err.contains("snd_pcm_poll_descriptors") || err.contains("alsa::poll() returned POLLERR") {
         "alsa_poll_failure"
@@ -3392,9 +3427,11 @@ async fn handle_audio_client(
 
     let history_replay_started_at = Instant::now();
 
-    // Serialize the entire history into a single in-memory blob so we can
-    // send it with one write_all + one flush, avoiding repeated partial
-    // flushes that dominate replay time for large histories.
+    // Build one in-memory blob of all history records, then split it into
+    // chunks at message boundaries before gzipping. A single oversized blob
+    // can exceed the 16 MiB MAX_HISTORY_PAYLOAD_SIZE the client enforces;
+    // chunking keeps each frame well under the limit while still amortising
+    // the gzip cost across many records per chunk.
     let history_blob = {
         // Estimate ~256 bytes per message; pre-allocate to avoid repeated
         // reallocation for large histories.
@@ -3478,24 +3515,34 @@ async fn handle_audio_client(
     };
     let (blob, replayed_history_count) = history_blob;
     if !blob.is_empty() {
-        // Gzip-compress the blob before sending. JSON history compresses very
-        // well (~10-20x) so this dramatically reduces both transfer size and
-        // the time the client spends waiting for data.
-        let compressed = {
-            let mut enc = GzEncoder::new(Vec::with_capacity(blob.len() / 8), Compression::fast());
-            enc.write_all(&blob)
+        // Target uncompressed bytes per chunk. JSON compresses ~10-20x so
+        // 8 MiB uncompressed lands well under the 16 MiB compressed cap even
+        // in adversarial cases (image-heavy LRPT/WEFAX history, low ratio).
+        const CHUNK_THRESHOLD_BYTES: usize = 8 * 1024 * 1024;
+
+        let mut total_compressed = 0usize;
+        let mut chunks_sent = 0usize;
+        for range in split_history_chunks(&blob, CHUNK_THRESHOLD_BYTES) {
+            let slice = &blob[range];
+            let mut enc = GzEncoder::new(Vec::with_capacity(slice.len() / 8), Compression::fast());
+            let compressed = enc
+                .write_all(slice)
                 .and_then(|_| enc.finish())
-                .unwrap_or(blob.clone())
-        };
+                .unwrap_or_else(|_| slice.to_vec());
             write_audio_msg(&mut writer, AUDIO_MSG_HISTORY_COMPRESSED, &compressed).await?;
+            total_compressed += compressed.len();
+            chunks_sent += 1;
+        }
+
         info!(
-            "Audio client {} replayed {} history messages in {:?} ({} → {} bytes, {:.1}x)",
+            "Audio client {} replayed {} history messages in {} chunk(s) in {:?} ({} → {} bytes, {:.1}x)",
             peer,
             replayed_history_count,
+            chunks_sent,
             history_replay_started_at.elapsed(),
             blob.len(),
-            compressed.len(),
-            blob.len() as f64 / compressed.len().max(1) as f64,
+            total_compressed,
+            blob.len() as f64 / total_compressed.max(1) as f64,
         );
     }