tts: seed the vocoder left context with the ICL reference tail

The upstream pipeline decodes reference plus generated codes then trims,
giving the first generated frames causal context from the reference; the
generated only decode started the vocoder from an empty state and
colored the clone onset. Seed both decode paths with the last
min(ref_T, left_ctx_frames) reference frames: the streaming decoder
takes them below its emit cursor so they are never emitted, the buffered
path prepends them and strips their samples. Raising
codec_left_context_sec past the reference duration reproduces the
upstream full reference decode exactly.

src/codec-chunked-decode.h

@@ -112,6 +112,26 @@ struct codec_chunked_decoder_stream {
         by_k.assign((size_t) K, {});
     }
 
+    // Seed the left context with the tail of the ICL reference codes so
+    // the first emitted chunk draws causal context from the reference
+    // instead of an empty decoder state, matching the upstream pipeline
+    // which decodes reference plus generated then trims. ref_kt is K
+    // major [K, ref_T]; the last min(ref_T, left_ctx_frames) frames are
+    // kept. Call once, after init and before any push_frame; the seeded
+    // frames sit below emit_start_frame so they are never emitted.
+    void seed_reference(const int32_t * ref_kt, int ref_T) {
+        int seed = ref_T < left_ctx_frames ? ref_T : left_ctx_frames;
+        if (seed <= 0) {
+            return;
+        }
+        for (int k = 0; k < K; k++) {
+            const int32_t * row = ref_kt + (size_t) k * (size_t) ref_T + (size_t) (ref_T - seed);
+            by_k[(size_t) k].insert(by_k[(size_t) k].end(), row, row + seed);
+        }
+        T_so_far         = seed;
+        emit_start_frame = seed;
+    }
+
     // Append one frame (K int32 codes, one per codebook). Drain any
     // chunks that became emittable. Returns false on decode failure or
     // when cb returns false (cancellation).

src/pipeline-tts.cpp

@@ -583,6 +583,11 @@ qt_status pipeline_tts_synthesize(PipelineTTS *                pt,
     codec_chunked_decoder_stream stream;
     if (streaming) {
         stream.init(num_codebooks, chunk_frames, left_ctx_frames);
+        // ICL clone: the reference tail seeds the decoder left context
+        // so the onset is voiced with the reference's causal state.
+        if (ref_codes_ptr != NULL) {
+            stream.seed_reference(ref_codes_ptr, ref_codes_T);
+        }
     }
 
     for (int step = 0; step < params->max_new_tokens; step++) {
@@ -794,23 +799,42 @@ qt_status pipeline_tts_synthesize(PipelineTTS *                pt,
     // equivalent to a single pipeline_codec_decode call when T_frames
     // fits in one chunk, bounded VRAM beyond that. Transpose codes from
     // [T_frames, K] to [K, T_frames] because codec_chunked_decode
-    // expects K major layout.
+    // expects K major layout. On the ICL clone path the tail of the
+    // reference codes prepends the buffer so the onset is voiced with
+    // the reference's causal state, mirroring the upstream pipeline
+    // which decodes reference plus generated then trims; the seeded
+    // samples strip from the front afterwards. Raising
+    // codec_left_context_sec past the reference duration reproduces the
+    // upstream full reference decode exactly.
     const int T_frames = (int) all_codes.size();
-    std::vector<int32_t> codes_kt((size_t) num_codebooks * (size_t) T_frames);
-    for (int t = 0; t < T_frames; t++) {
+    int       seed     = 0;
+    if (ref_codes_ptr != NULL) {
+        seed = ref_codes_T < left_ctx_frames ? ref_codes_T : left_ctx_frames;
+    }
+    const int            T_dec = seed + T_frames;
+    std::vector<int32_t> codes_kt((size_t) num_codebooks * (size_t) T_dec);
     for (int k = 0; k < num_codebooks; k++) {
-            codes_kt[(size_t) k * (size_t) T_frames + (size_t) t] = all_codes[(size_t) t][(size_t) k];
+        int32_t * row = codes_kt.data() + (size_t) k * (size_t) T_dec;
+        if (seed > 0) {
+            std::memcpy(row, ref_codes_ptr + (size_t) k * (size_t) ref_codes_T + (size_t) (ref_codes_T - seed),
+                        (size_t) seed * sizeof(int32_t));
+        }
+        for (int t = 0; t < T_frames; t++) {
+            row[(size_t) (seed + t)] = all_codes[(size_t) t][(size_t) k];
         }
     }
     Timer              t_codec;
     std::vector<float> audio =
-        codec_chunked_decode(&pt->codec, codes_kt.data(), num_codebooks, T_frames, chunk_frames, left_ctx_frames);
+        codec_chunked_decode(&pt->codec, codes_kt.data(), num_codebooks, T_dec, chunk_frames, left_ctx_frames);
     perf.codec_ms += t_codec.ms();
     if (audio.empty()) {
         qt_set_error("pipeline_tts_synthesize: codec decode returned no audio");
         qt_log(QT_LOG_ERROR, "[Pipeline] codec decode returned no audio");
         return QT_STATUS_GENERATE_FAILED;
     }
+    if (seed > 0) {
+        audio.erase(audio.begin(), audio.begin() + (size_t) seed * (size_t) TOKENIZER_HOP_LENGTH);
+    }
 
     if (params->dump_dir) {
         DebugDumper d;