[router] Apply chat template before cache-aware hashing (fix overlap=0 on chat traffic) (#27386)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

experimental/sgl-router/tests/e2e/infra/model_pool.py

@@ -26,7 +26,6 @@ import socket
 import subprocess
 import time
 from dataclasses import dataclass, field
-from pathlib import Path
 
 import httpx
 
@@ -34,16 +33,6 @@ from .model_specs import get_model_spec
 
 logger = logging.getLogger(__name__)
 
-# Passthrough Jinja chat template that emits ONLY `messages[*].content`
-# joined with `\n` — matching the router's cache_aware_zmq prompt
-# extraction. A worker launched with
-# ``--chat-template <PASSTHROUGH_CHAT_TEMPLATE_PATH>`` tokenizes the
-# raw content string, so its KV-block hashes align with what the
-# router computes from the same chat-completions request. Test-only.
-PASSTHROUGH_CHAT_TEMPLATE_PATH = str(
-    Path(__file__).parent / "passthrough_chat_template.jinja"
-)
-
 
 def _get_open_port() -> int:
     """Allocate an ephemeral TCP port in the range [20000, 55535].