fix(release): ggml_backend_is_cuda is unlinkable in the shipped .so; gate cosyvoice3's tail per backend

Two things, both from the v0.8.33 tag run.

1. THE RELEASE BUILD FAILED TO LINK. build-linux-x86_64-cuda and -cuda13 both
   died with:

       libcrispasr.so.0.8.33: undefined reference to `ggml_backend_is_cuda'

   Same class as the ggml_backend_is_cpu break fixed earlier this cycle:
   ggml_backend_is_cuda is a CUDA-MODULE symbol and is not linkable into the
   shared library. Two call sites had it — mimo_tokenizer.cpp and parakeet.cpp.

   Both now test the backend NAME via ggml_backend_name(), which is core ggml,
   matching what granite_speech, dots_tts and omnivoice already do — copying the
   code that demonstrably links rather than inventing a fourth approach. ROCm is
   included for the same reason granite_speech includes it. Swept src/*.cpp for
   the rest of the class: the only other hit, ggml_backend_is_metal, is a
   comment.

   Every other tag workflow passed — Docker, language wrappers, Win CUDA13,
   Piper Windows, ISA fallback and all four bindings — so this is the release
   artifacts only, not the tag.

2. CRISPASR_COSYVOICE3_CAMPP_TAIL=legacy, a per-backend switch for the
   unresolved tail-divisor question. Both paths synthesise correctly (8/8 each),
   so this is a CONSISTENCY control, not correctness: by default `--voice
   ref.wav` differs from the baked voice bank by cos ~0.998 for the same voice,
   because the shipped bank was produced with the old divisor. `legacy` makes
   the two agree.

   Backend-specific ON PURPOSE. CRISPASR_CAMPP_LEGACY_SEGPOOL already exists but
   is global, and would drag chatterbox, confucius4, dots-tts and fireredtts3
   away from their own settled PyTorch references to answer a question that is
   only open for cosyvoice3.

Author: CrispStrobe

docs/environment-variables.md

@@ -1603,3 +1603,21 @@ end-to-end cosine cannot do.
 - `CRISPASR_ZONOS_TTS_BENCH`
 - `CRISPASR_ZONOS_TTS_TEXT`
 - `CRISPASR_ZONOS_VULKAN_NATIVE`
+
+### `CRISPASR_COSYVOICE3_CAMPP_TAIL`
+
+`legacy` makes cosyvoice3's CAM++ speaker encoder use the old partial-tail
+divisor. Default is the same convention as every other CAM++ backend.
+
+Both paths synthesise correctly — the TTS→ASR roundtrip is 8/8 on each — so this
+is a CONSISTENCY switch, not a correctness one. The eight speaker embeddings
+baked into the shipped `cosyvoice3-voices.gguf` were produced with the old
+divisor, so by default a voice cloned from a WAV and the same voice taken from
+the bank differ by cos ~0.998. Set this to `legacy` if you need those two paths
+to agree.
+
+Why cosyvoice3 specifically is unsettled: its upstream is `campplus.onnx`, and
+two onnxruntime builds disagree about `AveragePool(ceil_mode=1)` on the same
+clip. `CRISPASR_CAMPP_LEGACY_SEGPOOL` also exists but is GLOBAL — it would drag
+chatterbox, confucius4, dots-tts and fireredtts3 away from their own settled
+PyTorch references to answer a cosyvoice3-only question.

src/cosyvoice3_tts.cpp

@@ -3121,6 +3121,33 @@ ggml_tensor* cv3_dit_block_apply(ggml_context* ctx0, ggml_tensor* x, ggml_tensor
 // runtime detector counts 133 per graph — q/k/v/o and both FFN layers across all
 // 22 blocks — far more than reading the source suggested. Gated; the fold is an
 // exact restatement (a linear is per-token independent). See core/quant_bcast.h.
+// #431-adj: which CAM++ tail divisor THIS backend uses, selectable.
+//
+// cosyvoice3's upstream is campplus.onnx and the question is genuinely open:
+// two onnxruntime builds disagree on AveragePool(ceil_mode=1) for the same
+// clip, while the eight embeddings baked into the shipped cosyvoice3-voices.gguf
+// match the LEGACY divisor exactly (cos 1.000000, |x| 14.1197 on zero_shot).
+//
+// Both paths synthesise correctly — the end-to-end roundtrip is 8/8 on each —
+// so this is not a correctness switch, it is a CONSISTENCY one:
+//
+//   default (fixed)  matches the other four CAM++ backends and the torch
+//                    reference; `--voice ref.wav` differs from the baked bank
+//                    by cos ~0.998 for the same voice.
+//   legacy           matches the baked voice bank, so a voice cloned from a WAV
+//                    and the same voice taken from the bank agree.
+//
+// CRISPASR_COSYVOICE3_CAMPP_TAIL=legacy selects the second. Backend-specific on
+// purpose: CRISPASR_CAMPP_LEGACY_SEGPOOL exists too, but it is global and would
+// drag chatterbox, confucius4, dots-tts and fireredtts3 away from their own
+// (settled) PyTorch references to fix a cosyvoice3-only question.
+static campplus_segpool::tail_divisor cosyvoice3_campp_tail() {
+    const char* e = crispasr_env::get("CRISPASR_COSYVOICE3_CAMPP_TAIL");
+    if (e && (std::strcmp(e, "legacy") == 0 || std::strcmp(e, "kernel") == 0))
+        return campplus_segpool::tail_divisor::kernel_size;
+    return campplus_segpool::tail_divisor::window_width;
+}
+
 static inline ggml_tensor* CV3MM(ggml_context* c, ggml_tensor* w, ggml_tensor* x) {
     // Default ON: verified on the shipped q4_k LLM + q8_0 flow — the detector
     // reports 133 broadcasting quantized matmuls per graph without the fold and
@@ -5620,7 +5647,8 @@ bool cv3_extract_native_runtime_voice(cosyvoice3_tts_context* ctx, const char* w
         return false;
 
     std::vector<float> native_spk =
-        chatterbox_campplus::embed_speaker(ctx->campplus.model, ctx->campplus.cache, pcm16.data(), (int)pcm16.size());
+        chatterbox_campplus::embed_speaker(ctx->campplus.model, ctx->campplus.cache, pcm16.data(), (int)pcm16.size(),
+                                           /*stats_var_floor=*/0.0f, cosyvoice3_campp_tail());
     if (native_spk.size() != 192)
         return false;
 
@@ -6353,7 +6381,8 @@ extern "C" int cosyvoice3_tts_extract_spk_emb(struct cosyvoice3_tts_context* ctx
         if (sr != 16000)
             pcm = core_audio::resample_polyphase(pcm.data(), (int)pcm.size(), sr, 16000);
         auto emb =
-            chatterbox_campplus::embed_speaker(ctx->campplus.model, ctx->campplus.cache, pcm.data(), (int)pcm.size());
+            chatterbox_campplus::embed_speaker(ctx->campplus.model, ctx->campplus.cache, pcm.data(), (int)pcm.size(),
+                                               /*stats_var_floor=*/0.0f, cosyvoice3_campp_tail());
         if (emb.size() != 192)
             return -1;
         std::memcpy(out_spk_emb, emb.data(), 192 * sizeof(float));

src/mimo_tokenizer.cpp

@@ -443,7 +443,14 @@ extern "C" struct mimo_tokenizer_context* mimo_tokenizer_init_from_file(const ch
     if (const char* e = std::getenv("CRISPASR_MIMO_TOK_CPU"); e && *e && *e != '0')
         weights_be = ctx->backend_cpu;
 #if defined(GGML_USE_CUDA)
-    ctx->cuda_rvq_available = weights_be == ctx->backend && ggml_backend_is_cuda(ctx->backend);
+    // ggml_backend_is_cuda() is a CUDA-MODULE symbol: linking it into
+    // libcrispasr.so fails with "undefined reference" in the release CUDA
+    // build. Test the backend NAME instead, which is core ggml and is what
+    // granite_speech, dots_tts and omnivoice already do. ROCm is included
+    // for the same reason granite_speech includes it.
+    const char* be_name = ctx->backend ? ggml_backend_name(ctx->backend) : nullptr;
+    ctx->cuda_rvq_available =
+        weights_be == ctx->backend && be_name && (std::strstr(be_name, "CUDA") || std::strstr(be_name, "ROCm"));
 #endif
     core_gguf::WeightLoad wl;
     if (!core_gguf::load_weights(path_model, weights_be, "mimo_tokenizer", wl)) {

src/parakeet.cpp

@@ -1404,7 +1404,13 @@ static bool parakeet_gpu_encoder_projection(const parakeet_context* ctx) {
     if (const char* e = crispasr_env::get("CRISPASR_RNNT_GPU_ENC_PROJ"))
         return *e == '1';
 #if defined(GGML_USE_CUDA)
-    return ggml_backend_is_cuda(ctx->backend);
+    // ggml_backend_is_cuda() is a CUDA-MODULE symbol: linking it into
+    // libcrispasr.so fails with "undefined reference" in the release CUDA
+    // build. Test the backend NAME instead, which is core ggml and is what
+    // granite_speech, dots_tts and omnivoice already do. ROCm is included
+    // for the same reason granite_speech includes it.
+    const char* be_name = ctx->backend ? ggml_backend_name(ctx->backend) : nullptr;
+    return be_name && (std::strstr(be_name, "CUDA") || std::strstr(be_name, "ROCm"));
 #else
     (void)ctx;
     return false;