fix(omnivoice): sync vendored diarizer reader to IFGO + fail-closed epoch guard

tools/omnivoice/CMakeLists.txt

@@ -396,6 +396,26 @@ target_include_directories(omnivoice-test-abi-c PRIVATE
     ${CMAKE_CURRENT_BINARY_DIR})
 target_link_libraries(omnivoice-test-abi-c PRIVATE omnivoice_lib)
 
+# omnivoice-test-diarizer-metadata : metadata-only GGUF contract test for the
+# vendored diarizer reader. It writes tiny GGUF files, so it does not need the
+# pyannote model artifact.
+add_executable(omnivoice-test-diarizer-metadata tests/diarizer-metadata.c)
+set_target_properties(omnivoice-test-diarizer-metadata PROPERTIES
+    C_STANDARD 99
+    C_STANDARD_REQUIRED ON
+    C_EXTENSIONS ON)
+if(MSVC)
+    target_compile_options(omnivoice-test-diarizer-metadata PRIVATE /W4 /WX)
+else()
+    target_compile_options(omnivoice-test-diarizer-metadata PRIVATE -Wall -Werror -pedantic)
+endif()
+target_include_directories(omnivoice-test-diarizer-metadata PRIVATE
+    ${OMNIVOICE_VOICE_CLASSIFIER_INCLUDE_DIRS})
+target_link_libraries(omnivoice-test-diarizer-metadata PRIVATE eliza_voice_classifiers)
+if(BUILD_TESTING)
+    add_test(NAME omnivoice-test-diarizer-metadata COMMAND omnivoice-test-diarizer-metadata)
+endif()
+
 # omnivoice-dac-parity : focused CPU numerical parity harness for
 # dac_conv_t1d's ggml_conv_transpose_1d migration (elizaOS/eliza#7660).
 # The default run uses deterministic synthetic DAC block shapes; pass

tools/omnivoice/src/voice-classifiers/voice_classifier/voice_diarizer.c

@@ -70,6 +70,8 @@
 #define DIAR_LINEAR0_OUT 128
 #define DIAR_LINEAR1_OUT 128
 #define DIAR_LEAKY_ALPHA 0.01f
+#define DIAR_GGUF_CONVERTER_EPOCH 2
+#define DIAR_LSTM_GATE_ORDER "IFGO"
 
 /* Cached pointers + buffer struct for one diarizer session. */
 struct voice_diarizer_session {
@@ -223,12 +225,11 @@ static void lstm_run_dir(const float *x_dot_W,
             gate_buf[g] = acc;
         }
 
-        /* Apply nonlinearities. Gate order I, O, F, C (ONNX IOFC layout —
-         * pyannote-3 segmentation GGUF, #9460). */
+        /* Apply nonlinearities. Gate order I, F, G, O. */
         const float *gi = gate_buf + 0 * H;
-        const float *go = gate_buf + 1 * H;
-        const float *gf = gate_buf + 2 * H;
-        const float *gg = gate_buf + 3 * H;
+        const float *gf = gate_buf + 1 * H;
+        const float *gg = gate_buf + 2 * H;
+        const float *go = gate_buf + 3 * H;
         for (int j = 0; j < H; ++j) {
             const float i_t = sigmoidf(gi[j]);
             const float f_t = sigmoidf(gf[j]);
@@ -393,6 +394,33 @@ int voice_diarizer_open(const char *gguf, voice_diarizer_handle *out) {
         meta.sample_rate != VOICE_CLASSIFIER_SAMPLE_RATE_HZ) return -EINVAL;
     if (meta.num_classes != 0 &&
         meta.num_classes != VOICE_DIARIZER_NUM_CLASSES) return -EINVAL;
+    if (meta.window_samples != 0 &&
+        meta.window_samples != DIAR_WINDOW_SAMPLES) return -EINVAL;
+    if (meta.frames_per_window != 0 &&
+        meta.frames_per_window != DIAR_FRAMES_PER_WINDOW) return -EINVAL;
+    if (meta.lstm_layers != 0 &&
+        meta.lstm_layers != DIAR_LSTM_LAYERS) return -EINVAL;
+    if (meta.lstm_hidden != 0 &&
+        meta.lstm_hidden != DIAR_LSTM_HIDDEN) return -EINVAL;
+    if (meta.linear0_out != 0 &&
+        meta.linear0_out != DIAR_LINEAR0_OUT) return -EINVAL;
+    if (meta.linear1_out != 0 &&
+        meta.linear1_out != DIAR_LINEAR1_OUT) return -EINVAL;
+    if (meta.converter_epoch < DIAR_GGUF_CONVERTER_EPOCH) {
+        fprintf(stderr,
+                "[voice_diarizer] stale GGUF converter epoch %d; need >= %d with LSTM gates packed as %s\n",
+                meta.converter_epoch,
+                DIAR_GGUF_CONVERTER_EPOCH,
+                DIAR_LSTM_GATE_ORDER);
+        return -EINVAL;
+    }
+    if (strcmp(meta.lstm_gate_order, DIAR_LSTM_GATE_ORDER) != 0) {
+        fprintf(stderr,
+                "[voice_diarizer] unsupported LSTM gate order '%s'; expected %s\n",
+                meta.lstm_gate_order[0] ? meta.lstm_gate_order : "<missing>",
+                DIAR_LSTM_GATE_ORDER);
+        return -EINVAL;
+    }
 
     struct voice_diarizer_session *s =
         (struct voice_diarizer_session *)calloc(1, sizeof(*s));

tools/omnivoice/src/voice-classifiers/voice_classifier/voice_gguf_loader.c

@@ -226,6 +226,17 @@ static int vc_gguf_load_state_cb(const char *key,
         free(str);
         return 1;
     }
+    if (vc_gguf_key_eq(key, s->want_prefix, "lstm_gate_order")) {
+        if (type != VC_GGUF_TYPE_STRING) return -1;
+        char *str = NULL;
+        const int rc = vc_gguf_read_string(f, &str);
+        if (rc != 0) return -1;
+        const size_t n = sizeof(s->out->lstm_gate_order) - 1;
+        strncpy(s->out->lstm_gate_order, str, n);
+        s->out->lstm_gate_order[n] = '\0';
+        free(str);
+        return 1;
+    }
     /* Wav2Small-specific uint32 keys: read into the matching field
      * and let the per-head opener validate against its expectation. */
     struct {
@@ -239,6 +250,13 @@ static int vc_gguf_load_state_cb(const char *key,
         { "ffn_dim",    &s->out->ffn_dim    },
         { "num_layers", &s->out->num_layers },
         { "num_heads",  &s->out->num_heads  },
+        { "converter_epoch",   &s->out->converter_epoch   },
+        { "window_samples",    &s->out->window_samples    },
+        { "frames_per_window", &s->out->frames_per_window },
+        { "lstm_layers",       &s->out->lstm_layers       },
+        { "lstm_hidden",       &s->out->lstm_hidden       },
+        { "linear0_out",       &s->out->linear0_out       },
+        { "linear1_out",       &s->out->linear1_out       },
     };
     for (size_t i = 0; i < sizeof(u32_extras) / sizeof(u32_extras[0]); ++i) {
         if (vc_gguf_key_eq(key, s->want_prefix, u32_extras[i].suffix)) {

tools/omnivoice/src/voice-classifiers/voice_classifier/voice_gguf_loader.h

@@ -57,6 +57,18 @@ typedef struct voice_gguf_metadata {
     int ffn_dim;
     int num_layers;
     int num_heads;
+    /* Diarizer-specific metadata. Older published GGUFs leave these
+     * zero/empty; the diarizer opener treats missing gate metadata as legacy
+     * and validates explicit metadata against the reader's compiled contract.
+     */
+    int converter_epoch;
+    int window_samples;
+    int frames_per_window;
+    int lstm_layers;
+    int lstm_hidden;
+    int linear0_out;
+    int linear1_out;
+    char lstm_gate_order[16];
 } voice_gguf_metadata_t;
 
 /* GGUF GGML tensor data types — we only support F32 in the per-head