feat(cli): filter chat-template control tokens from print_token

The --chat flag wraps the prompt with chat templates (Gemma 4, Phi-3,
Llama 3, ChatML). The model then sometimes emits those template tokens
back into the output stream, which the CLI printed literally as garbage.

Now filters the same set as the server:
  <|think|>  <think>  </think>  <|channel>  <|turn>  <turn|>
  <|end|>  <|assistant|>  <|user|>  <|system|>
  <|im_end|>  <|im_start|>
  <start_of_turn>  <end_of_turn>
  <|begin_of_text|>  <|end_of_text|>
  <|start_header_id|>  <|end_header_id|>  <|eot_id|>

Before: Gemma 4 E2B --chat "Calculate 2+2..." → garbage "ussererererer..."
After:  Gemma 4 E2B --chat "Calculate 2+2..." → "The number is **4**."

All 35 unit tests + 7 regression tests pass.

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

Author: unamedkr

tools/quant.c

@@ -62,9 +62,27 @@ static int clock_gettime(int id, struct timespec* ts) {
 /* Forward-pass profiling flag (defined in tq_transformer.c) */
 extern int g_tq_profile_enabled;
 
-/* Streaming token callback */
+/* Streaming token callback — filters chat-template control tokens that
+ * would otherwise leak into CLI output when --chat is active. */
 static void print_token(const char* text, void* user_data) {
     (void)user_data;
+    if (!text || !text[0]) return;
+
+    /* Skip thinking / template tokens (same list as server). Gemma 4 raw
+     * output contains <|think|>, Qwen3 uses <think>/</think>, chat-templated
+     * models may emit <|end|>/<|im_end|>/<|eot_id|> etc. */
+    if (strstr(text, "<|think|>") || strstr(text, "<think>") ||
+        strstr(text, "</think>") || strstr(text, "<|channel>") ||
+        strstr(text, "<|turn>") || strstr(text, "<turn|>") ||
+        strstr(text, "<|end|>") || strstr(text, "<|assistant|>") ||
+        strstr(text, "<|user|>") || strstr(text, "<|system|>") ||
+        strstr(text, "<|im_end|>") || strstr(text, "<|im_start|>") ||
+        strstr(text, "<start_of_turn>") || strstr(text, "<end_of_turn>") ||
+        strstr(text, "<|begin_of_text|>") || strstr(text, "<|end_of_text|>") ||
+        strstr(text, "<|start_header_id|>") || strstr(text, "<|end_header_id|>") ||
+        strstr(text, "<|eot_id|>"))
+        return;
+
     fputs(text, stdout);
     fflush(stdout);
 }