chore: cleanup garbage code (#29770)

experimental/sgl-router/BENCHMARKS.md

@@ -32,12 +32,12 @@ cargo bench --bench radix_tree_benchmark -- --sample-size 30 --measurement-time
 cargo bench --bench manual_policy_benchmark
 ```
 
-For the quick smoke runs whose numbers are reproduced below: drop
+For the quick runs whose numbers are reproduced below: drop
 `--sample-size` to 10 and `--measurement-time` to 2 (Criterion will
 warn about reduced statistical confidence but the order-of-magnitude
 comparison stands).
 
-## Smoke-run data points (M1 MacBook, release profile)
+## Quick-run Data Points (M1 MacBook, release profile)
 
 These are NOT the real acceptance numbers — they're a sanity check
 that the sgl-router routing primitives are in the same ballpark as the

experimental/sgl-router/src/workers/manager.rs

@@ -680,7 +680,7 @@ mod tests {
         let _ = manager_handle.await;
     }
 
-    /// End-to-end wiring smoke test: spin up a fake worker, run the
+    /// End-to-end wiring check: spin up a fake worker, run the
     /// manager with a real `KvEventIndex` against that worker URL, and
     /// verify both `Added` and `Removed` propagate through to the
     /// index's internal worker map.

experimental/sgl-router/tests/e2e/chat_completions/test_validation.py

@@ -7,7 +7,7 @@ The shape:
   - non-streaming + streaming chat completion
   - assistant message non-empty, role correct, finish_reason set
 
-These are the smoke tests that run first; if they pass, the heavier
+These quick checks run first; if they pass, the heavier
 multi-worker acceptance tests are worth running.
 """
 

experimental/sgl-router/tests/e2e/conftest.py

@@ -2,7 +2,7 @@
 
 Two flavors of fixtures coexist here:
 
-  1. **Session-scoped smoke fixtures** (``sglang_server`` + ``router``) —
+  1. **Session-scoped sanity fixtures** (``sglang_server`` + ``router``) —
      launch ONE SGLang worker + ONE router on fixed ports for the whole
      test session. Used by the lightweight ``test_chat_smoke.py`` /
      ``test_tokenize_smoke.py`` files. These are the cheap "did the
@@ -174,7 +174,7 @@ def build_smoke_router_args(
     tokenizer_path: str,
     sglang_url: str,
 ) -> list[str]:
-    """Build the sgl-router CLI flags the smoke ``router`` fixture launches.
+    """Build the sgl-router CLI flags the single-worker ``router`` fixture launches.
 
     Static single-worker discovery (``--worker-urls``) pointed at the one
     SGLang worker, serving exactly one model.

experimental/sgl-router/tests/e2e/pyproject.toml

@@ -4,7 +4,7 @@
 
 [tool.pytest.ini_options]
 minversion = "8.0"
-# Default discovery: smoke tests (top-level test_*.py) and the
+# Default discovery: top-level test_*.py files and the
 # multi-worker chat_completions suite. k8s_integration is intentionally
 # not in the default set — it requires a kind/k8s cluster and is
 # invoked explicitly.

experimental/sgl-router/tests/e2e/requirements.txt

@@ -9,7 +9,7 @@ pytest-asyncio==0.24.0
 # pin of `huggingface_hub==0.26.2` here got installed AFTER the SGLang
 # deps and downgraded huggingface_hub past `is_offline_mode`'s top-
 # level export, which broke `from sglang.srt.server_args import …`
-# at module import time and turned every smoke test into a 5-minute
+# at module import time and turned every quick check into a 5-minute
 # `/health` timeout with no actionable signal.
 # The e2e suite only uses huggingface_hub's `try_to_load_from_cache`,
 # which is available in every release SGLang would install.

experimental/sgl-router/tests/proxy/pd_pool_isolation.rs

@@ -13,7 +13,7 @@
 //!   (existing code path; pinned here so a future PD wiring change
 //!   doesn't silently swap codes).
 //! * A PD-disagg model with both pools healthy → request flows to the
-//!   prefill worker (smoke; the decode worker MUST NOT be selected for
+//!   prefill worker (sanity check; the decode worker MUST NOT be selected for
 //!   the chat route).
 
 use axum::body::Body;