feat(breeze_tts): sub-chunk streaming via resumable AR stepper

include/engine/models/breeze_tts/generator.h

@@ -29,6 +29,12 @@ struct BreezeGenerationRequest {
     uint64_t seed = 0;
 };
 
+struct BreezeStreamStep {
+    std::vector<int32_t> new_codes;
+    int64_t new_frames = 0;
+    bool done = false;
+};
+
 class BreezeGeneratorRuntime {
 public:
     BreezeGeneratorRuntime(
@@ -43,6 +49,16 @@ class BreezeGeneratorRuntime {
     engine::runtime::AudioBuffer generate(const BreezeGenerationRequest & request);
     BreezeSpeechCodes encode_reference(const engine::runtime::AudioBuffer & audio) const;
 
+    // Resumable sub-chunk streaming. begin_stream runs prompts + prefills;
+    // step_stream runs up to max_new_frames AR frames; end_stream releases
+    // graphs. Only one stream may be active at a time. decode_codes runs the
+    // codec over a code prefix without releasing graphs (caller brackets with
+    // begin/end_stream).
+    void begin_stream(const BreezeGenerationRequest & request);
+    BreezeStreamStep step_stream(size_t max_new_frames);
+    void end_stream();
+    engine::runtime::AudioBuffer decode_codes(const std::vector<int32_t> & codes, int64_t frames);
+
 private:
     struct Impl;
     std::unique_ptr<Impl> impl_;

include/engine/models/breeze_tts/session.h

@@ -74,11 +74,20 @@ class BreezeTTSSession final
     std::unique_ptr<BreezeGeneratorRuntime> generator_;
     engine::runtime::CacheSlots<ReferenceCacheKey, ReferenceCacheEntry, ReferenceCacheKeyEqual> reference_cache_;
     std::optional<ReferenceCacheEntry> uncached_reference_;
+    std::optional<engine::runtime::StreamEvent> next_subchunk_event();
     std::vector<engine::runtime::TaskRequest> stream_chunk_requests_;
     std::optional<BreezeSpeechCodes> stream_reference_codes_;
     engine::runtime::AudioBuffer stream_merged_audio_;
     size_t stream_chunk_index_ = 0;
     bool stream_started_ = false;
+    bool stream_subchunk_ = false;
+    size_t stream_frames_per_event_ = 32;
+    int64_t stream_lookahead_margin_ = 12;
+    bool stream_chunk_active_ = false;
+    std::vector<int32_t> stream_codes_;
+    int64_t stream_total_frames_ = 0;
+    size_t stream_emitted_samples_ = 0;
+    size_t stream_event_seq_ = 0;
 };
 
 }  // namespace engine::models::breeze_tts

model_specs/breeze_tts.json

@@ -127,6 +127,29 @@
         "required": false,
         "min": 0,
         "default": 0
+      },
+      {
+        "name": "stream_subchunk",
+        "type": "bool",
+        "description": "Emit sub-chunk streaming events (multiple audio deltas per text chunk) instead of one event per chunk.",
+        "required": false,
+        "default": false
+      },
+      {
+        "name": "stream_frames_per_event",
+        "type": "int",
+        "description": "Acoustic frames generated per sub-chunk streaming event.",
+        "required": false,
+        "min": 1,
+        "default": 32
+      },
+      {
+        "name": "stream_lookahead_margin",
+        "type": "int",
+        "description": "Trailing codec frames withheld from emission to absorb convolutional boundary artifacts.",
+        "required": false,
+        "min": 0,
+        "default": 12
       }
     ],
     "session": [