config: retire the alias-form process-global config reads

`sa = get_server_args()` followed by `sa.field` reads the same startup record as
the direct form; the read ratchet added in the previous slice pinned twelve of
them as the remaining surface. Eleven now read the accessor for what they
actually want:

- `is_enable_moe_cp_allgather` compares the attention-CP and MoE-DP sizes to
  decide whether a forward needs an allgather, so it reads the live topology
  through `get_parallel()` — the same source `get_moe_cp_size()` right above it
  already uses. Both groups exist once model-parallel init has run, which is
  before any forward.
- The DeepSeek MLA decode-backend gate and Inkling's attention paths read
  `get_exec().kernel`; Inkling's KV-dtype checks read `get_model()`. These are
  per-runner fields, and the value they get is the config published for the
  runner being built — unchanged from what the alias returned.
- The int8 mamba checkpoint pool reads `get_exec().mamba`. It keeps its guard
  for callers that construct the pool with no published config; that guard now
  catches the namespace accessor instead of the slot.

`model_loader`'s `moe_dp_size` stays on the instance and is exempt: the dict it
belongs to already reports the live size under `"dp"`, so that entry is the
configured intent, and `get_parallel()` shadows the name with the live value.

Alias-form baseline 12 -> 0. What remains on `get_server_args()` in the package
is the derived API (properties and methods computed from several fields plus the
HF config) and four config-intent reads of live-shadowed sizes, each exempt by
name with its reason.

test/registered/unit/models/test_kimi_k3_vision.py

@@ -21,7 +21,7 @@ from sglang.srt.multimodal.kimi_k3_vit_cuda_graph_runner import (
     KimiK3ViTCudaGraphRunner,
 )
 from sglang.srt.multimodal.mm_utils import run_dp_sharded_mrope_vision_model
-from sglang.srt.runtime_context import get_parallel
+from sglang.srt.runtime_context import get_context, get_parallel
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=1, suite="base-a-test-cpu")
@@ -357,7 +357,7 @@ def test_kimi_k3_position_interpolation_uses_contiguous_chw(monkeypatch):
     assert torch.equal(actual, expected)
 
 
-def test_kimi_k3_prepares_shared_attention_metadata_once(monkeypatch):
+def test_kimi_k3_prepares_shared_attention_metadata_once(monkeypatch, request):
     metadata_ids = []
     values_are_contiguous = []
 
@@ -370,11 +370,13 @@ def test_kimi_k3_prepares_shared_attention_metadata_once(monkeypatch):
             values_are_contiguous.append(v.is_contiguous())
             return q
 
-    monkeypatch.setattr(
-        kimi_k3_vl,
-        "get_server_args",
-        lambda: SimpleNamespace(mm_attention_backend="flashinfer_cudnn"),
+    # Force the backend through the context, not by patching an import binding:
+    # production reads the published config bag (get_mm().mm_attention_backend).
+    override = get_context().override_server_args(
+        mm_attention_backend="flashinfer_cudnn"
     )
+    override.install()
+    request.addfinalizer(override.restore)
     monkeypatch.setitem(kimi_k3_vl.QKV_BACKEND_IMPL, "flashinfer_cudnn", FakeAttention)
 
     encoder = MoonViT3dEncoder(

test/registered/unit/test_global_config_read_ratchet.py

@@ -31,6 +31,12 @@ What legitimately remains:
     reads ``get_dcp_group()``, and that group is only installed when DCP is on.
   - ``cuda_ipc_transport_utils.tp_size`` runs in the tokenizer process, which has
     no groups at all (the call site already guards for "not published yet").
+  - ``dp_attention.attn_cp_size`` / ``moe_dp_size``: the configuration the
+    predicate detects (``attn_cp_size > moe_dp_size``) is the one where
+    ``initialize_model_parallel`` aliases ``_MOE_DP`` to ``_ATTN_CP``, so the live
+    sizes are equal there and a live comparison is always false.
+  - ``model_loader/loader.py`` reports both: the same dict carries the live
+    ``moe_dp_size`` under ``"dp"``, so this entry is the configured intent.
 - The alias-form baseline is not zero yet. Lowering it is the next slice; the
   failure message lists the sites whenever the count moves.
 """
@@ -67,13 +73,16 @@ _DERIVED_MEMBERS = frozenset(
 _CONFIG_INTENT_SIZES = frozenset(
     {
         ("srt/layers/attention/dsa/dsa_indexer.py", "pp_size"),
+        ("srt/layers/dp_attention.py", "attn_cp_size"),
+        ("srt/layers/dp_attention.py", "moe_dp_size"),
         ("srt/mem_cache/allocation.py", "dcp_size"),
+        ("srt/model_loader/loader.py", "moe_dp_size"),
         ("srt/utils/cuda_ipc_transport_utils.py", "tp_size"),
     }
 )
 
 _DIRECT_BASELINE = 0
-_ALIAS_BASELINE = 12
+_ALIAS_BASELINE = 0
 
 
 def _is_global_call(node) -> bool: