tts-cpp: chatterbox-mtl — q8 KV-cache e2e coverage in test_multilingual_synth

Adds end-to-end regression coverage for the q8-on-GPU align-cast fix, on top of
the op-level test_metal_ops sentinel:

  - test_multilingual_synth: new --kv-cache-type passthrough (+ CHATTERBOX_KV_CACHE_TYPE
    env fallback), forwarded to the CLI. Lets the existing MTL synth harness —
    which already synthesizes on the GPU, writes a WAV, and validates audio
    sanity (RMS/peak/clipping/silence) — exercise the quantized KV path.

  - CMake: register mtl-synth-q8-<lang> for a diverse script subset
    (en/ar/ru/hi = Latin/Arabic/Cyrillic/Devanagari), labelled `mtl-q8`. On a
    Metal fleet these hit the exact path that used to SIGABRT ("unsupported op
    'CONT'") before the fix; the script diversity doubles as cross-language
    alignment coverage under quantization. Same env model wiring
    (CHATTERBOX_T3_MTL / CHATTERBOX_S3GEN) as the f32 mtl-synth-* tests.

Local validation (stock ggml Metal, M2) via an Engine-level WAV harness: q8 on
Metal produces real audio across en/es/fr/de/ar/ru/hi/ko (no crash, finite,
non-clipped); q8-vs-f32 alignment tracks closely (e.g. ar q8 5.12s/rms .051 vs
f32 5.00s/rms .051).

Refs QVAC-19557

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: ogad-tether

tts-cpp/CMakeLists.txt

@@ -745,6 +745,41 @@ if (TTS_CPP_BUILD_TESTS)
         endif()
     endforeach()
 
+    # QVAC-19557: q8_0 KV-cache regression for the MTL variant.  A quantized KV
+    # cache used to SIGABRT on Metal ("unsupported op 'CONT'") via the alignment
+    # probe; the dequantizing-cast fix recovers q8 KV on the GPU.  These run the
+    # full synth with --kv-cache-type q8_0 on whatever backend the fleet has
+    # (Metal fleets exercise the original crash path) and validate the WAV is
+    # real audio (RMS/peak/clipping/duration) — i.e. an end-to-end regression
+    # the op-level test_metal_ops sentinel can't cover.  A diverse script subset
+    # (Latin/Arabic/Cyrillic/Devanagari) doubles as cross-language alignment
+    # coverage under quantization.  Labelled `mtl-q8` so a Metal fleet can select
+    # them; same env model wiring (CHATTERBOX_T3_MTL / CHATTERBOX_S3GEN) as the
+    # f32 mtl-synth-* tests above.
+    set(_mtl_q8_phrases
+        "en|Hello, this is a multilingual text-to-speech test."
+        "ar|مرحبًا، هذا اختبار متعدد اللغات لتحويل النص إلى كلام."
+        "ru|Привет, это многоязычный тест синтеза речи."
+        "hi|नमस्ते, यह एक बहुभाषी वाक् संश्लेषण परीक्षण है।"
+    )
+    foreach(_entry IN LISTS _mtl_q8_phrases)
+        string(REPLACE "|" ";" _parts "${_entry}")
+        list(GET _parts 0 _lang)
+        list(GET _parts 1 _text)
+        add_test(
+            NAME mtl-synth-q8-${_lang}
+            COMMAND $<TARGET_FILE:test-multilingual-synth>
+                --lang           "${_lang}"
+                --text           "${_text}"
+                --kv-cache-type  q8_0
+                --out            "${_mtl_out_dir}/${_lang}-q8.wav"
+        )
+        set_tests_properties(mtl-synth-q8-${_lang} PROPERTIES
+            LABELS "multilingual;mtl-q8"
+            TIMEOUT 180
+        )
+    endforeach()
+
     # End-to-end EOS round-trip regression.  Drives tts-cli to
     # synthesize a set of English phrases, transcribes with whisper-cli, and
     # asserts the transcription is close to the input (CER guard -> catches

tts-cpp/test/test_multilingual_synth.cpp

@@ -170,6 +170,7 @@ void usage(const char * prog) {
 struct Args {
     std::string t3_path, s3gen_path, lang, text, out_path;
     std::string mecab_dict, cangjie_tsv;
+    std::string kv_cache_type;   // "" -> CLI default (f32); "q8_0"/"f16" exercise the quantized KV path
     int seed = 42;
     int n_gpu_layers = 99;
     bool verbose = false;
@@ -189,6 +190,7 @@ bool parse_args(int argc, char ** argv, Args & args) {
         else if (a == "--out")          { auto v = next("--out");          if (!v) return false; args.out_path = v; }
         else if (a == "--seed")         { auto v = next("--seed");         if (!v) return false; args.seed = std::atoi(v); }
         else if (a == "--n-gpu-layers") { auto v = next("--n-gpu-layers"); if (!v) return false; args.n_gpu_layers = std::atoi(v); }
+        else if (a == "--kv-cache-type") { auto v = next("--kv-cache-type"); if (!v) return false; args.kv_cache_type = v; }
         else if (a == "--verbose" || a == "-v") { args.verbose = true; }
         else if (a == "--mecab-dict")  { auto v = next("--mecab-dict");  if (!v) return false; args.mecab_dict = v; }
         else if (a == "--cangjie-tsv") { auto v = next("--cangjie-tsv"); if (!v) return false; args.cangjie_tsv = v; }
@@ -215,6 +217,10 @@ void resolve_env_fallbacks(Args & args) {
         const char * env = std::getenv("CHATTERBOX_CANGJIE_TSV");
         if (env && *env) args.cangjie_tsv = env;
     }
+    if (args.kv_cache_type.empty()) {
+        const char * env = std::getenv("CHATTERBOX_KV_CACHE_TYPE");
+        if (env && *env) args.kv_cache_type = env;
+    }
 }
 
 int check_language_registry(const std::string & lang) {
@@ -243,6 +249,10 @@ int run_synthesis(const Args & args) {
         "--n-gpu-layers", std::to_string(args.n_gpu_layers),
     };
     if (args.verbose) cli_args.push_back("--verbose");
+    if (!args.kv_cache_type.empty()) {
+        cli_args.push_back("--kv-cache-type");
+        cli_args.push_back(args.kv_cache_type);
+    }
     if (!args.mecab_dict.empty()) {
         cli_args.push_back("--mecab-dict");
         cli_args.push_back(args.mecab_dict);