nit: drop the dead flow_temp hparam, the pack table holds the default

conversion/pockettts.py

@@ -169,7 +169,7 @@ class PocketTTSMmprojModel(MmprojModel):
         self.gguf_writer.add_audio_num_mel_bins(1)
 
         # generation: flow-matching decoder + mimi decoder
-        # note: the SEANet and flow net hparams are hardcoded on the clip.cpp side for now
+        # the SEANet and flow net hparams are constant across the family, clip.cpp holds them
         self.gguf_writer.add_clip_has_gen_audio_encoder(True)
         self.gguf_writer.add_clip_gen_audio_projector_type(gguf.VisionProjectorType.POCKETTTS_GEN)
         self.gguf_writer.add_gen_audio_projection_dim(self.n_embd_text)

tools/mtmd/clip-model.h

@@ -148,7 +148,6 @@ struct clip_hparams {
     int32_t mimi_downsample   = 0;      // encoder frame rate / model frame rate
     int32_t mimi_tfm_context  = 0;      // attention window of the mimi transformers, in frames
     int32_t flow_n_step       = 1;      // lsd_decode steps
-    float   flow_temp         = 0.7f;   // noise std is sqrt(temp), the caller can override it
 
     // qwen3tts code2wav
     int32_t wav_tfm_n_layer      = 0;

tools/mtmd/clip.cpp

@@ -4851,7 +4851,8 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
                 } else {
                     // flow matching starts from gaussian noise, std = sqrt(temp)
                     ggml_tensor * t = get_inp_tensor("inp_noise");
-                    const float temp = params->flow_temp > 0.0f ? params->flow_temp : hparams.flow_temp;
+                    // Config.default_temperature, a caller that knows its variant overrides it
+                    const float temp = params->flow_temp > 0.0f ? params->flow_temp : 0.7f;
                     std::normal_distribution<float> dist(0.0f, std::sqrt(temp));
                     std::vector<float> noise(ggml_nelements(t));
                     for (auto & v : noise) {