fix(gemma4): PLE buffer overrun broke Gemma 4 E4B

The per-layer embedding (PLE) stack buffers were sized 8960 floats,
matching E2B exactly (35 layers × 256 ple_dim). E4B has 42 layers,
requiring 10752 elements — a 1792-float stack overrun that caused
the forward pass to hang indefinitely (stack corruption triggering
infinite loop somewhere downstream).

Fixed in both src/engine/tq_transformer.c and quant.h:
- temp_embd and temp_proj buffers bumped to 16384 floats
- Added safety guard for total_ple > 16384

Before: E4B Q8_0 prompt=5 tokens generates 0 tokens in 15 minutes
        (stuck in stack-corruption-induced hang)
After:  E4B Q8_0 generates "4. This is a mathematical fact."
        E4B Q4_0 generates "4。请问，您能用更"

All 35 unit tests pass, all 7 model regression tests pass.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: unamedkr

quant.h

@@ -15481,10 +15481,11 @@ float* tq_forward(tq_model_t* model, tq_state_t* s, int token, int pos) {
             s->ple_buf = (float*)calloc((size_t)total_ple, sizeof(float));
         }
 
-        /* Step A: Dequant per_layer_token_embd[token] → temp_embd[8960]
+        /* Step A: Dequant per_layer_token_embd[token] → temp_embd
          * The embedding tensor is [total_ple, vocab_size] in GGUF row-major,
          * so one token's data is at row offset = token * row_bytes. */
-        float temp_embd[8960];  /* stack buffer, total_ple <= 8960 */
+        float temp_embd[16384];  /* stack buffer, E2B=8960, E4B=10752 */
+        if (total_ple > 16384) return s->logits;  /* safety guard */
         {
             size_t type_size = tq_ggml_type_size(model->ple_embedding_type);
             int blck = tq_ggml_type_blck(model->ple_embedding_type);
@@ -15500,11 +15501,11 @@ float* tq_forward(tq_model_t* model, tq_state_t* s, int token, int pos) {
             temp_embd[i] *= ple_scale;
         }
 
-        /* Step B: per_layer_model_proj @ embed_raw → temp_proj[8960]
+        /* Step B: per_layer_model_proj @ embed_raw → temp_proj
          * ple_proj is [total_ple, hidden_dim] FP32 (rows=8960, cols=1536).
          * We need: for each output row d in [0, total_ple): dot(ple_proj[d,:], s->x[:])
          * Note: s->x already has the scaled embedding from above. */
-        float temp_proj[8960];
+        float temp_proj[16384];
         tq_matmul(temp_proj, s->x, model->ple_proj, total_ple, dim);
 
         /* Scale by 1/sqrt(hidden_dim) */

src/engine/tq_transformer.c

@@ -2342,17 +2342,18 @@ float* tq_forward(tq_model_t* model, tq_state_t* s, int token, int pos) {
     if (model->ple_dim > 0 && model->ple_embedding && model->ple_proj && !getenv("TQ_NO_PLE")) {
         int ple_dim = model->ple_dim;
         int n_layers = c->n_layers;
-        int total_ple = n_layers * ple_dim;  /* e.g., 35 * 256 = 8960 */
+        int total_ple = n_layers * ple_dim;  /* E2B: 35*256=8960, E4B: 42*256=10752 */
 
         /* Lazy allocation of ple_buf */
         if (!s->ple_buf) {
             s->ple_buf = (float*)calloc((size_t)total_ple, sizeof(float));
         }
 
-        /* Step A: Dequant per_layer_token_embd[token] → temp_embd[8960]
+        /* Step A: Dequant per_layer_token_embd[token] → temp_embd
          * The embedding tensor is [total_ple, vocab_size] in GGUF row-major,
          * so one token's data is at row offset = token * row_bytes. */
-        float temp_embd[8960];  /* stack buffer, total_ple <= 8960 */
+        float temp_embd[16384];  /* stack buffer, sized for total_ple up to 16384 (E4B=10752) */
+        if (total_ple > 16384) return s->logits;  /* safety guard — should not happen */
         {
             size_t type_size = tq_ggml_type_size(model->ple_embedding_type);
             int blck = tq_ggml_type_blck(model->ple_embedding_type);
@@ -2368,11 +2369,11 @@ float* tq_forward(tq_model_t* model, tq_state_t* s, int token, int pos) {
             temp_embd[i] *= ple_scale;
         }
 
-        /* Step B: per_layer_model_proj @ embed_raw → temp_proj[8960]
-         * ple_proj is [total_ple, hidden_dim] FP32 (rows=8960, cols=1536).
+        /* Step B: per_layer_model_proj @ embed_raw → temp_proj
+         * ple_proj is [total_ple, hidden_dim] FP32. For E4B: [10752, 2560].
          * We need: for each output row d in [0, total_ple): dot(ple_proj[d,:], s->x[:])
          * Note: s->x already has the scaled embedding from above. */
-        float temp_proj[8960];
+        float temp_proj[16384];
         tq_matmul(temp_proj, s->x, model->ple_proj, total_ple, dim);
 
         /* Scale by 1/sqrt(hidden_dim) */