diff --git a/.claude/rules/general-code-style.md b/.claude/rules/general-code-style.md
index dc09944d8..07f82c548 100644
--- a/.claude/rules/general-code-style.md
+++ b/.claude/rules/general-code-style.md
@@ -9,7 +9,7 @@ Default conventions for new and modified Python code. Prefer these unless there
 
 - **Prefer stateless.** Favor pure functions over methods that mutate instance state; pass inputs in, return outputs out.
 - **Prefer immutable.** Default to immutable data (frozen structs, tuples, read-only values); mutate only when there is a clear need.
-- **Extract init-static values at construction.** When a derived value's inputs are frozen for the object's lifetime (typically configuration: constructor args, env vars, server args), compute it once in `__init__` and store it as a well-named attribute (`self.mtp_enabled`, `self.needs_cpu_seq_lens`); later code reads the attribute instead of re-deriving it. Input immutability is the hard precondition — if inputs can change, recompute in place or funnel mutation through a single override point (the frozen `ServerArgs.override()` pattern). If you can't give the value a meaningful name, the boundary is wrong — don't cache unnameable subexpressions.
+- **Extract init-static values at construction.** When a derived value's inputs are frozen for the object's lifetime (typically configuration: constructor args, env vars, server args), compute it once in `__init__` and store it as a well-named attribute (`self.mtp_enabled`, `self.needs_cpu_seq_lens`); later code reads the attribute instead of re-deriving it. Input immutability is the hard precondition — if inputs can change, recompute in place or funnel mutation through a single override point (`get_context().override()` for resolved config). If you can't give the value a meaningful name, the boundary is wrong — don't cache unnameable subexpressions.
 - **Functions stay small.** Keep each function under ~100 LOC; split larger ones into named helpers.
 - **Files stay small.** Keep each file under ~2k LOC; split larger modules along cohesive boundaries.
 - **Core functions read like pseudocode.** The main / orchestration function of a unit should be short and read like algorithm pseudocode — push detail into well-named helpers so the top-level flow is obvious.
diff --git a/.claude/skills/sglang-runtime-context/SKILL.md b/.claude/skills/sglang-runtime-context/SKILL.md
index a5f68174a..a5690abcd 100644
--- a/.claude/skills/sglang-runtime-context/SKILL.md
+++ b/.claude/skills/sglang-runtime-context/SKILL.md
@@ -49,12 +49,41 @@ resolved configuration lives in the namespace bags.**
   `get_context().override(source, **fields)`. It writes the bag leaves in place
   (namespace readers see the new value) and records provenance in the overrides log.
   There is **no write-through** to the `ServerArgs` instance — it stays pristine.
-  `ServerArgs.override(...)` (instance-only) is being retired; its call sites are
-  ratcheted down and new ones are rejected.
-- **Nested publishes**: a construction step that must publish a private copy (the
-  draft-worker build publishes the draft's rewritten config for the duration of the
-  build) wraps itself in `get_context().preserve_config()` — the enclosing lifecycle,
-  including its post-publish overrides, is value-snapshotted and reinstated on exit.
+  There is no in-place mutation entry on the instance at all: it is read-only after
+  resolution.
+- **Late launcher-stage resolution (pre-publish)**: a few rules cannot run inside
+  `__post_init__` — LoRA normalization, and the auto-parser detection that needs a
+  tokenizer/chat-template load. They are resolution, not mutation, and they write
+  **in place** via `arg_groups.overrides.declare_late_resolution(server_args,
+  source, **fields)`, which refuses the published instance. In place is the point:
+  every holder of that object must see the resolved value — the HTTP server, the
+  multi-tokenizer workers it is serialized for, the schedulers it forks. Returning a
+  variant here is a bug: the launcher rebinds its local and everyone else keeps the
+  unresolved object.
+- **A config another runner / worker / process is built from**: `server_args.derive(
+  source, **fields)` returns a variant (an encode worker's `base_gpu_id`/`tp_size`).
+  The receiver — and any bags projected from it — are untouched; resolution does
+  **not** re-run, so do not reach for it to "re-resolve" a config.
+- **Per-runner values inside one process are constructor arguments, not a variant.**
+  The draft worker's `context_length`, load format and attention backend travel as
+  arguments to `TpModelWorker` / `ModelRunner` and live on the runner
+  (`ModelRunner.draft_attention_backend`, `kv_cache_dtype_str`, …), because target
+  and draft coexist and the process-wide bags can only describe one of them.
+
+**Why a bag override cannot stand in for the last two.** The bags are projected at
+publish *from the instance's fields*, so anything the runtime must read has to be on
+the instance before publish — an override afterwards puts instance and bags back out
+of agreement, and whole-object readers (`ModelConfig.from_server_args`,
+`build_load_config`, `MMEncoder`'s own `self.server_args.X`) never see it. And bags do
+not cross a process boundary: a child publishes from the object it receives and
+re-projects its own bags, so a parent-side override is lost. Values that feed
+construction before any bag exists (group init reads `server_args.tp_size`) have no
+bag to override at all.
+
+- **Nested publishes**: a construction step that must publish a private copy wraps
+  itself in `get_context().preserve_config()` — the enclosing lifecycle, including its
+  post-publish overrides, is value-snapshotted and reinstated on exit. The draft build
+  no longer needs it: per-runner values are constructor arguments now.
 
 ### Reads that legitimately stay on a `ServerArgs` instance
 
@@ -209,18 +238,18 @@ ONE thread — do not design for TBO threads that don't exist.
 ## Guardrails (these fail CI; what to do when they fire)
 
 1. **Strict mutation guard** (always on): bare `server_args.x = ...` after resolution
-   raises unconditionally — `ServerArgs.__setattr__` no longer consults
-   `SGLANG_STRICT_CONFIG_MUTATION` (the env var survives only as a legacy harness
-   flag). Projected bags are sealed the same way (leaf assignment raises — write via
-   `get_context().override`).
+   raises unconditionally in `ServerArgs.__setattr__` — this *is* the guarantee that
+   no writer can desync the bags, so there is no writer ratchet any more. Change
+   resolved config with `get_context().override`, build a per-runner config with
+   `server_args.derive`. Projected bags are sealed the same way (leaf assignment
+   raises).
 2. **Mutation ratchet** (`test_server_args_mutation_ratchet.py`, exact pin 0 over the whole
    package minus the pipeline / multimodal_gen): textual scan for assignment forms. Never
    raise the baseline.
-3. **Writer ratchet** (`test_server_args_writer_ratchet.py`): `ServerArgs.override`
-   call sites are pinned exactly and may only shrink — instance writes never reach the
-   bags, so namespace readers desync from the writer. New post-publish writes go through
-   `get_context().override`; rerouting a writer means flipping **all its readers to the
-   bag in the same commit** (no transitional dual-write).
+3. **Derive contract** (`test_server_args_derive.py`): deriving leaves the receiver
+   intact, a published config still refuses assignment, and deriving does not publish.
+   Rerouting a writer to the bags means flipping **all its readers in the same commit**
+   (no transitional dual-write).
 4. **Legacy-accessor ratchet** (`test_legacy_global_ratchet.py`): `get_global_server_args`
    call sites must not grow — new code uses `runtime_context.get_server_args()` (and
    business decisions should read the bags).
diff --git a/python/sglang/compile_deep_gemm.py b/python/sglang/compile_deep_gemm.py
index cf733c79e..7eb283d2c 100644
--- a/python/sglang/compile_deep_gemm.py
+++ b/python/sglang/compile_deep_gemm.py
@@ -169,21 +169,18 @@ def launch_server_process_and_send_one_request(
     )
 
 
-def refine_server_args(server_args: ServerArgs, compile_args: CompileArgs):
-    # Disable cuda graph and torch compile to save time. Writes after
-    # ServerArgs.__post_init__ don't propagate to cuda_graph_config via the
-    # legacy disable_cuda_graph field, so flip both phases directly.
+def compile_server_args(args, compile_args: CompileArgs) -> ServerArgs:
+    """The config this script serves with: no cuda graph, no torch compile, and a
+    watchdog that outlives the compilation."""
+    args.enable_torch_compile = False
+    # Watchdog timeout follows compile_args.timeout because compilation takes long.
+    args.watchdog_timeout = compile_args.timeout
+    args.warmups = "compile-deep-gemm"
+    server_args = ServerArgs.from_cli_args(args)
     server_args.cuda_graph_config[Phase.DECODE].backend = Backend.DISABLED
     server_args.cuda_graph_config[Phase.PREFILL].backend = Backend.DISABLED
     print(f"Disable CUDA Graph and Torch Compile to save time...")
-
-    # Watchdog timeout follows compile_args.timeout because compilation takes long.
-    server_args.override(
-        "compile_deep_gemm.refine_server_args",
-        enable_torch_compile=False,
-        watchdog_timeout=compile_args.timeout,
-        warmups="compile-deep-gemm",
-    )
+    return server_args
 
 
 def run_compile(server_args: ServerArgs, compile_args: CompileArgs):
@@ -216,9 +213,7 @@ if __name__ == "__main__":
     ServerArgs.add_cli_args(parser)
     CompileArgs.add_cli_args(parser)
     args = parser.parse_args()
-    server_args = ServerArgs.from_cli_args(args)
     compile_args = CompileArgs.from_cli_args(args)
-
-    refine_server_args(server_args, compile_args)
+    server_args = compile_server_args(args, compile_args)
 
     run_compile(server_args, compile_args)
diff --git a/python/sglang/lang/backend/runtime_endpoint.py b/python/sglang/lang/backend/runtime_endpoint.py
index fa5414afd..9b453f72f 100644
--- a/python/sglang/lang/backend/runtime_endpoint.py
+++ b/python/sglang/lang/backend/runtime_endpoint.py
@@ -384,13 +384,15 @@ class Runtime:
         from sglang.srt.server_args import ServerArgs
         from sglang.srt.utils.network import is_port_available
 
-        self.server_args = ServerArgs(*args, log_level=log_level, **kwargs)
-
-        # Pre-allocate ports
-        for port in range(self.server_args.port, 40000):
+        # Pre-allocate a port before building the config, so the config is born
+        # with the port this runtime will serve on.
+        requested_port = kwargs.pop(
+            "port", ServerArgs.__dataclass_fields__["port"].default
+        )
+        for port in range(requested_port, 40000):
             if is_port_available(port):
                 break
-        self.server_args.override("runtime_endpoint.port_alloc", port=port)
+        self.server_args = ServerArgs(*args, log_level=log_level, port=port, **kwargs)
 
         self.url = self.server_args.url()
         self.generate_url = self.url + "/generate"
diff --git a/python/sglang/srt/arg_groups/overrides.py b/python/sglang/srt/arg_groups/overrides.py
index 245f5943c..f5af1ddbc 100644
--- a/python/sglang/srt/arg_groups/overrides.py
+++ b/python/sglang/srt/arg_groups/overrides.py
@@ -208,12 +208,47 @@ def run_post_process_pass(server_args: Any, fn: Callable[..., dict]) -> None:
 def _apply_fields(server_args: Any, fields: Dict[str, Any]) -> None:
     """Write fields on behalf of the pipeline (bypasses the strict bare-
     assignment guard that protects post-resolution mutation)."""
-    object.__setattr__(server_args, "_in_override", True)
+    object.__setattr__(server_args, "_internal_write", True)
     try:
         for field, value in fields.items():
             setattr(server_args, field, value)
     finally:
-        object.__setattr__(server_args, "_in_override", False)
+        object.__setattr__(server_args, "_internal_write", False)
+
+
+def declare_late_resolution(server_args: Any, source: str, **fields: Any) -> None:
+    """Resolve fields on a config that is **not published yet**.
+
+    A few resolution rules cannot run inside ``__post_init__``: LoRA
+    normalization and the auto-parser detection need the launcher's validation
+    stage (and, for the parsers, a tokenizer / chat-template load). They still
+    belong to the resolution pipeline — they decide what the process will run
+    with — so they write the fields in place, before anything publishes the
+    object. Writing in place is the point: every holder of that instance (the
+    HTTP server, the multi-tokenizer workers it serializes for, the schedulers
+    it forks) must see the resolved value.
+
+    Refuses to touch the published instance: after publish the bags exist and a
+    field write would desync them, which is what ``get_context().override`` is
+    for.
+    """
+    from sglang.srt.runtime_context import get_context
+
+    try:
+        published = get_context().server_args
+    except ValueError:
+        published = None
+    if published is server_args:
+        raise ValueError(
+            f"declare_late_resolution({source!r}) called on the published config; "
+            "post-publish changes go to the bags via get_context().override(...)"
+        )
+    log = getattr(server_args, "_runtime_mutations", None)
+    if log is None:
+        log = []
+        object.__setattr__(server_args, "_runtime_mutations", log)
+    log.append((source, dict(fields)))
+    _apply_fields(server_args, fields)
 
 
 def materialize_declarations(server_args: Any) -> None:
diff --git a/python/sglang/srt/disaggregation/encode_server.py b/python/sglang/srt/disaggregation/encode_server.py
index 84732b837..600942960 100644
--- a/python/sglang/srt/disaggregation/encode_server.py
+++ b/python/sglang/srt/disaggregation/encode_server.py
@@ -1,7 +1,6 @@
 import asyncio
 import concurrent.futures
 import contextlib
-import copy
 import ctypes
 import functools
 import logging
@@ -3520,10 +3519,7 @@ async def run_dp_worker(
     # gpu_id is the device chosen by maybe_reindex_device_id in the parent:
     # 0 when CVD is pinned to one GPU, else the absolute id. rank=0, so
     # MMEncoder runs set_device(base_gpu_id).
-    args = copy.deepcopy(server_args)
-    # The copy is already resolved (read-only); route the per-worker
-    # specialization through the audited mutation entry.
-    args.override("encode_server.dp_worker", base_gpu_id=gpu_id, tp_size=1)
+    args = server_args.derive("encode_server.dp_worker", base_gpu_id=gpu_id, tp_size=1)
     enc = MMEncoder(args, dist_init_method=f"tcp://127.0.0.1:{get_free_port()}", rank=0)
 
     global encoder_metrics_collector
diff --git a/python/sglang/srt/environ.py b/python/sglang/srt/environ.py
index eee06bdc2..92421fae1 100644
--- a/python/sglang/srt/environ.py
+++ b/python/sglang/srt/environ.py
@@ -247,10 +247,6 @@ class DsparkFoldedSampling(IntEnum):
 
 class Envs:
 
-    # Raise on bare server_args field assignments after resolution; mutation
-    # must go through ServerArgs.override() (enabled by the test harness).
-    SGLANG_STRICT_CONFIG_MUTATION = EnvBool(False)
-
     # Per-role config-namespace bookkeeping: off / record / enforce (value is
     # validated fail-loud in runtime_context, which resolves it once at import
     # so the read stays dynamo-prunable).
diff --git a/python/sglang/srt/mem_cache/hi_mamba_radix_cache.py b/python/sglang/srt/mem_cache/hi_mamba_radix_cache.py
index 8671a7a75..fb8815b70 100644
--- a/python/sglang/srt/mem_cache/hi_mamba_radix_cache.py
+++ b/python/sglang/srt/mem_cache/hi_mamba_radix_cache.py
@@ -99,16 +99,6 @@ class HiMambaRadixCache(MambaRadixCache):
 
     def __init__(self, params: CacheInitParams, server_args: ServerArgs):
         self._enable_metrics_flag = params.enable_metrics
-        if server_args.hicache_io_backend == "direct":
-            if server_args.hicache_mem_layout == "page_first":
-                server_args.override(
-                    "hicache.mem_layout_force", hicache_mem_layout="page_first_direct"
-                )
-                logger.warning(
-                    "Page first layout is not supported with direct IO backend, "
-                    "switching to page first direct layout"
-                )
-
         self.page_size = params.page_size
         self.hybrid_kv_cache = params.token_to_kv_pool_allocator.get_kvcache()
         if not isinstance(self.hybrid_kv_cache, HybridLinearKVPool):
diff --git a/python/sglang/srt/parser/template_detection.py b/python/sglang/srt/parser/template_detection.py
index d61c4d3ef..d353aa1c6 100644
--- a/python/sglang/srt/parser/template_detection.py
+++ b/python/sglang/srt/parser/template_detection.py
@@ -22,13 +22,15 @@ import logging
 import os
 import re
 from dataclasses import dataclass
-from typing import Callable, Optional, Tuple
+from typing import Callable, Dict, Optional, Tuple
 
 import jinja2
 import jinja2.ext
 import jinja2.nodes
 import jinja2.sandbox
 
+from sglang.srt.arg_groups.overrides import declare_late_resolution
+
 logger = logging.getLogger(__name__)
 
 
@@ -625,26 +627,21 @@ def detect_inline_system_support(chat_template: Optional[str]) -> bool:
         return False
 
 
-def _resolve_auto_parser(
-    server_args,
+def _detect_auto_parser(
     attr: str,
     ctx: TemplateDetectionContext,
     rules: Tuple[DetectionRule, ...],
     label: str,
-) -> None:
-    """Resolve a single auto parser, updating server_args in place."""
+) -> Optional[str]:
+    """The parser one auto field resolves to (``None`` disables it)."""
     detected = match_rules(ctx, rules, label)
     if detected:
-        server_args.override(source="template-detection", **{attr: detected})
         logger.info(
             f"Auto-detected --{attr.replace('_', '-')} as '{detected}' from chat template"
         )
-    else:
-        logger.warning(
-            f"--{attr.replace('_', '-')}=auto specified but could not detect "
-            f"{label} from chat template. Disabling {label}."
-        )
-        server_args.override(source="template-detection", **{attr: None})
+        return detected
+    _log_undetected_parser(attr, label)
+    return None
 
 
 def _load_explicit_jinja_template(chat_template_arg: Optional[str]) -> Optional[str]:
@@ -658,15 +655,15 @@ def _load_explicit_jinja_template(chat_template_arg: Optional[str]) -> Optional[
         return f.read().replace("\\n", "\n")
 
 
-def _disable_auto_parser(server_args, attr: str, label: str) -> None:
+def _log_undetected_parser(attr: str, label: str) -> None:
     logger.warning(
         f"--{attr.replace('_', '-')}=auto specified but could not detect "
         f"{label} from chat template. Disabling {label}."
     )
-    server_args.override(source="template-detection", **{attr: None})
 
 
-def _resolve_architecture_auto_parsers(server_args) -> None:
+def _architecture_auto_parsers(server_args, needs: Tuple[str, ...]) -> Dict[str, str]:
+    """The parsers the model architecture implies, for the fields still on auto."""
     from sglang.srt.utils.hf_transformers_utils import get_config
 
     config = get_config(
@@ -686,30 +683,37 @@ def _resolve_architecture_auto_parsers(server_args) -> None:
     elif "DeepseekV3" in arch:
         reasoning_parser, tool_call_parser = "deepseek-v3", "deepseekv32"
     else:
-        return
+        return {}
 
+    resolved = {}
     for attr, detected in (
         ("reasoning_parser", reasoning_parser),
         ("tool_call_parser", tool_call_parser),
     ):
-        if getattr(server_args, attr) == "auto":
-            server_args.override(source="template-detection", **{attr: detected})
+        if attr in needs:
+            resolved[attr] = detected
             logger.info(
                 f"Auto-detected --{attr.replace('_', '-')} as '{detected}' "
                 f"from model architecture '{arch}'"
             )
+    return resolved
 
 
 def resolve_auto_parsers(server_args) -> None:
-    """Resolve --reasoning-parser=auto and --tool-call-parser=auto before scheduler.
+    """Resolve ``--reasoning-parser=auto`` / ``--tool-call-parser=auto`` from the
+    chat template, in place, before anything publishes ``server_args``.
 
-    This performs a lightweight tokenizer load to detect parsers from the chat
-    template. Called early in engine init before scheduler subprocesses are spawned.
+    Performs a lightweight tokenizer load, so it runs once in engine init. In
+    place because everyone who holds this instance must see the resolved value:
+    the schedulers it forks, the HTTP server, and the tokenizer workers it is
+    serialized for.
     """
-    needs_reasoning = server_args.reasoning_parser == "auto"
-    needs_tool_call = server_args.tool_call_parser == "auto"
-
-    if not needs_reasoning and not needs_tool_call:
+    needs = tuple(
+        attr
+        for attr in ("reasoning_parser", "tool_call_parser")
+        if getattr(server_args, attr) == "auto"
+    )
+    if not needs:
         return
 
     from sglang.srt.utils.hf_transformers_utils import get_tokenizer
@@ -741,6 +745,8 @@ def resolve_auto_parsers(server_args) -> None:
     ctx = build_detection_context(
         template, tokenizer, reasoning_config, force_reasoning
     )
+
+    detected: Dict[str, Optional[str]] = {}
     if ctx is None:
         if has_explicit_template_without_detection:
             logger.warning(
@@ -750,38 +756,26 @@ def resolve_auto_parsers(server_args) -> None:
             )
         else:
             try:
-                _resolve_architecture_auto_parsers(server_args)
+                detected.update(_architecture_auto_parsers(server_args, needs))
             except Exception as e:
                 logger.warning(
                     "Failed to load model config for architecture-based auto-detection: %s",
                     e,
                 )
-        if needs_reasoning:
-            if server_args.reasoning_parser == "auto":
-                _disable_auto_parser(
-                    server_args, "reasoning_parser", "reasoning parser"
-                )
-        if needs_tool_call:
-            if server_args.tool_call_parser == "auto":
-                _disable_auto_parser(
-                    server_args, "tool_call_parser", "tool-call parser"
-                )
-        return
+        for attr, label in (
+            ("reasoning_parser", "reasoning parser"),
+            ("tool_call_parser", "tool-call parser"),
+        ):
+            if attr in needs and attr not in detected:
+                _log_undetected_parser(attr, label)
+                detected[attr] = None
+    else:
+        for attr, rules, label in (
+            ("reasoning_parser", REASONING_PARSER_RULES, "reasoning parser"),
+            ("tool_call_parser", TOOL_CALL_PARSER_RULES, "tool-call parser"),
+        ):
+            if attr in needs:
+                detected[attr] = _detect_auto_parser(attr, ctx, rules, label)
 
-    if needs_reasoning:
-        _resolve_auto_parser(
-            server_args,
-            "reasoning_parser",
-            ctx,
-            REASONING_PARSER_RULES,
-            "reasoning parser",
-        )
-
-    if needs_tool_call:
-        _resolve_auto_parser(
-            server_args,
-            "tool_call_parser",
-            ctx,
-            TOOL_CALL_PARSER_RULES,
-            "tool-call parser",
-        )
+    if detected:
+        declare_late_resolution(server_args, "template-detection", **detected)
diff --git a/python/sglang/srt/runtime_context.py b/python/sglang/srt/runtime_context.py
index 33cd4780b..f958b3d3e 100644
--- a/python/sglang/srt/runtime_context.py
+++ b/python/sglang/srt/runtime_context.py
@@ -1011,8 +1011,7 @@ class _ServerArgsOverride:
 
     def install(self) -> ServerArgs:
         """Publish a fresh dummy-boundary ``ServerArgs`` carrying the
-        overrides (written through ``ServerArgs.override`` for provenance);
-        returns the published instance."""
+        overrides; returns the published instance."""
         from sglang.srt.server_args import ServerArgs
 
         assert not self._installed, "override_server_args already installed"
@@ -1025,7 +1024,7 @@ class _ServerArgsOverride:
         self._prev_capture = ctx.flags.capture.enable_torch_compile
         server_args = ServerArgs(model_path="dummy")
         if self._fields:
-            server_args.override(source="test-override", **self._fields)
+            server_args = server_args.derive("test-override", **self._fields)
         # The dummy boundary skips materialization, which would leave the
         # strict mutation guard unarmed on the published object — mark it
         # materialized so bare post-publish writes raise like they do on a
diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index a40d9607b..99bdfda2f 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -16,6 +16,7 @@
 from __future__ import annotations
 
 import argparse
+import copy
 import dataclasses
 import glob
 import importlib
@@ -8547,56 +8548,73 @@ class ServerArgs:
 
         return resolved_view(self)
 
-    def override(self, source: str, **fields) -> None:
-        """The single post-resolution mutation point.
+    def _late_resolution(self, source: str, **fields) -> None:
+        """Resolve fields at the launcher's validation stage (pre-publish).
 
-        After ``__post_init__`` the configuration is resolved; the audited
-        runtime adjustments (load-resolved values, control-plane
-        reconfiguration, deployment wiring) go through here instead of
-        assigning fields. Whitelisted resolvable fields also join the
-        declaration stash, so a republish resolves the same values;
-        everything is recorded with its ``source`` for provenance.
+        See ``arg_groups.overrides.declare_late_resolution``: in place, because
+        every holder of this instance must see the resolved value, and refused
+        outright once the config is published.
+        """
+        from sglang.srt.arg_groups.overrides import declare_late_resolution
+
+        declare_late_resolution(self, source, **fields)
+
+    def derive(self, source: str, **fields) -> ServerArgs:
+        """A copy carrying variant values: a draft worker's context length, an
+        encode worker's device, a launcher's late port pick.
+
+        The receiver is untouched, so a config already published from it -- and
+        the namespace bags projected out of it -- stay true; the variant is a
+        second config, to be published in its own right or handed to whoever
+        owns it. Resolution does not re-run: the values being set are decided
+        *after* it, from inputs resolution never had, and re-resolving a
+        resolved config re-derives conditional decisions from the wrong ones.
+
+        Whitelisted resolvable fields also join the copy's declaration stash so
+        a later re-resolution keeps them; ``source`` is recorded for provenance.
         """
         from sglang.srt.arg_groups.arg_utils import resolvable_fields
 
-        whitelist = resolvable_fields(type(self))
+        variant = copy.deepcopy(self)
+        whitelist = resolvable_fields(type(variant))
         declared = {k: v for k, v in fields.items() if k in whitelist}
         rest = {k: v for k, v in fields.items() if k not in whitelist}
         if declared:
-            stash = getattr(self, "_resolved_overrides", None)
+            stash = getattr(variant, "_resolved_overrides", None)
             if stash is None:
                 stash = []
-                object.__setattr__(self, "_resolved_overrides", stash)
+                object.__setattr__(variant, "_resolved_overrides", stash)
             stash.append((source, dict(declared)))
         if rest:
-            log = getattr(self, "_runtime_mutations", None)
+            log = getattr(variant, "_runtime_mutations", None)
             if log is None:
                 log = []
-                object.__setattr__(self, "_runtime_mutations", log)
+                object.__setattr__(variant, "_runtime_mutations", log)
             log.append((source, dict(rest)))
-        object.__setattr__(self, "_in_override", True)
+        object.__setattr__(variant, "_internal_write", True)
         try:
             for field, value in fields.items():
-                setattr(self, field, value)
+                setattr(variant, field, value)
         finally:
-            object.__setattr__(self, "_in_override", False)
+            object.__setattr__(variant, "_internal_write", False)
+        return variant
 
     def __setattr__(self, name, value):
-        # after materialization the fields are the resolved startup
-        # configuration -- the pristine, READ-ONLY record. A bare assignment
-        # outside ServerArgs.override() (and the resolution pipeline, which runs
-        # before materialization) always raises; resolved config is mutated on
-        # the context bags via get_context().override(...), not here. (Formerly
-        # gated on SGLANG_STRICT_CONFIG_MUTATION; now unconditional.)
+        # After materialization the fields are the resolved startup
+        # configuration -- the pristine, READ-ONLY record that the config bags
+        # were projected from. Resolved config changes go to the bags via
+        # get_context().override(source, ...); a config that differs per runner
+        # or per worker is a separate object, built with derive().
         if (
             not name.startswith("_")
             and getattr(self, "_declarations_materialized", False)
-            and not getattr(self, "_in_override", False)
+            and not getattr(self, "_internal_write", False)
         ):
             raise AttributeError(
                 f"server_args.{name} assigned after resolution; server_args is "
                 "read-only -- use get_context().override(source, ...) to change "
-                "resolved config."
+                "resolved config, or server_args.derive(source, ...) to build a "
+                "variant for one runner."
             )
         object.__setattr__(self, name, value)
 
@@ -8914,7 +8932,7 @@ class ServerArgs:
         # Enable LoRA if any LoRA paths are provided for backward compatibility.
         if self.lora_paths:
             if self.enable_lora is None:
-                self.override("check_lora_server_args", enable_lora=True)
+                self._late_resolution("check_lora_server_args", enable_lora=True)
                 logger.warning(
                     "--enable-lora is set to True because --lora-paths is provided."
                 )
@@ -8925,7 +8943,7 @@ class ServerArgs:
 
         if self.enable_lora:
             if self.enable_lora_overlap_loading is None:
-                self.override(
+                self._late_resolution(
                     "check_lora_server_args", enable_lora_overlap_loading=False
                 )
 
@@ -8984,9 +9002,11 @@ class ServerArgs:
                             "Expected a string or a dictionary."
                         )
                     parsed_lora_paths.append(lora_ref)
-                self.override("check_lora_server_args", lora_paths=parsed_lora_paths)
+                self._late_resolution(
+                    "check_lora_server_args", lora_paths=parsed_lora_paths
+                )
             elif isinstance(self.lora_paths, dict):
-                self.override(
+                self._late_resolution(
                     "check_lora_server_args",
                     lora_paths=[
                         LoRARef(
@@ -8999,7 +9019,7 @@ class ServerArgs:
                     ],
                 )
             elif self.lora_paths is None:
-                self.override("check_lora_server_args", lora_paths=[])
+                self._late_resolution("check_lora_server_args", lora_paths=[])
             else:
                 raise ValueError(
                     f"Invalid type for --lora-paths: {type(self.lora_paths)}. "
@@ -9009,7 +9029,7 @@ class ServerArgs:
             # Normalize target modules to a set; keep {"all"} as a sentinel
             # that gets resolved model-awarely in lora_manager.init_lora_shapes().
             if self.lora_target_modules:
-                self.override(
+                self._late_resolution(
                     "check_lora_server_args",
                     lora_target_modules=set(self.lora_target_modules),
                 )
diff --git a/python/sglang/srt/speculative/draft_worker_common.py b/python/sglang/srt/speculative/draft_worker_common.py
index 2e1685da1..adc23c728 100644
--- a/python/sglang/srt/speculative/draft_worker_common.py
+++ b/python/sglang/srt/speculative/draft_worker_common.py
@@ -1,7 +1,6 @@
 from __future__ import annotations
 
 import logging
-from copy import deepcopy
 from typing import TYPE_CHECKING, Optional
 
 import msgspec
@@ -43,11 +42,11 @@ class DraftWorkerBundle(msgspec.Struct, frozen=True):
 
 
 def _resolve_draft_attention_backend_fallback(
-    *, draft_server_args: ServerArgs, algo_label: str
+    *, server_args: ServerArgs, algo_label: str
 ) -> str:
-    draft_backend = draft_server_args.speculative_draft_attention_backend
+    draft_backend = server_args.speculative_draft_attention_backend
     if draft_backend is None:
-        draft_backend, _ = draft_server_args.get_attention_backends()
+        draft_backend, _ = server_args.get_attention_backends()
     if draft_backend is None:
         return "triton" if torch.version.hip else "flashinfer"
     if draft_backend not in _SUPPORTED_DRAFT_BACKENDS:
@@ -111,8 +110,7 @@ def draft_server_args_copy(server_args: ServerArgs, target_model_config) -> Serv
     for _source, fields in get_context().overrides_log():
         resolved.update(fields)
 
-    draft_server_args = deepcopy(server_args)
-    draft_server_args.override(
+    return server_args.derive(
         "draft_worker.copy",
         **{
             **resolved,
@@ -120,7 +118,6 @@ def draft_server_args_copy(server_args: ServerArgs, target_model_config) -> Serv
             **_draft_load_format_fields(),
         },
     )
-    return draft_server_args
 
 
 def build_draft_tp_worker(
@@ -133,18 +130,15 @@ def build_draft_tp_worker(
     algo_label: str,
     attention_backend_override: Optional[str] = None,
 ) -> DraftWorkerBundle:
-    draft_server_args = deepcopy(server_args)
     # An override names a draft-specific backend the caller has already
     # validated (e.g. a self-drafting architecture); it skips the generic
     # supported-backend fallback below.
     draft_backend = attention_backend_override or (
         _resolve_draft_attention_backend_fallback(
-            draft_server_args=draft_server_args, algo_label=algo_label
+            server_args=server_args, algo_label=algo_label
         )
     )
-    # Post-resolution ServerArgs rejects bare assignment; route the draft-copy
-    # adjustments through the audited mutation point.
-    draft_server_args.override(
+    draft_server_args = server_args.derive(
         "draft_worker.build",
         **draft_server_args_overrides(target_model_config, draft_backend),
     )
diff --git a/python/sglang/test/kits/attention_unittest/runner_modes/speculative_draft_runner.py b/python/sglang/test/kits/attention_unittest/runner_modes/speculative_draft_runner.py
index eb150e7c9..46db9aebf 100644
--- a/python/sglang/test/kits/attention_unittest/runner_modes/speculative_draft_runner.py
+++ b/python/sglang/test/kits/attention_unittest/runner_modes/speculative_draft_runner.py
@@ -295,7 +295,6 @@ def _configure_runner_for_eagle_draft(
     *,
     speculative_attention_mode: str = "decode",
 ) -> None:
-    server_args = runner.server_args
     updates = {
         "attention_backend": case.backend,
         "cuda_graph_config": CudaGraphConfig(
@@ -326,11 +325,12 @@ def _configure_runner_for_eagle_draft(
         "torch_compile_max_bs": 0,
         "use_mla_backend": runner.use_mla_backend,
     }
-    server_args.override(source="attention-unittest-eagle-draft", **updates)
-    # Re-publish so the bags pick up the overrides.
     from sglang.srt.runtime_context import get_context
 
-    get_context().set_server_args(server_args)
+    runner.server_args = runner.server_args.derive(
+        "attention-unittest-eagle-draft", **updates
+    )
+    get_context().set_server_args(runner.server_args)
 
     runner.spec_algorithm = SpeculativeAlgorithm.EAGLE
     runner.is_draft_worker = True
@@ -391,11 +391,11 @@ def _build_frozen_kv_mtp_fixture(
         runner_batch_size=settings.capture_batch_size,
     )
     _configure_runner_for_eagle_draft(fixture.runner, case, settings)
-    fixture.runner.server_args.override(
-        "attention_unittest.frozen_kv_draft", speculative_algorithm="FROZEN_KV_MTP"
-    )
     from sglang.srt.runtime_context import get_context
 
+    fixture.runner.server_args = fixture.runner.server_args.derive(
+        "attention_unittest.frozen_kv_draft", speculative_algorithm="FROZEN_KV_MTP"
+    )
     get_context().set_server_args(fixture.runner.server_args)
     fixture.runner.spec_algorithm = SpeculativeAlgorithm.FROZEN_KV_MTP
     fixture.runner.draft_attn_backend = fixture.backend
diff --git a/python/sglang/test/test_utils.py b/python/sglang/test/test_utils.py
index efd41ed4e..2a9c1a68d 100644
--- a/python/sglang/test/test_utils.py
+++ b/python/sglang/test/test_utils.py
@@ -9,10 +9,6 @@ import inspect
 import json
 import logging
 import os
-
-# Registered tests run with the strict config-mutation guard: bare
-# server_args assignments after resolution raise (use ServerArgs.override).
-os.environ.setdefault("SGLANG_STRICT_CONFIG_MUTATION", "1")
 import random
 import re
 import shlex
diff --git a/test/registered/attention/unittests/hybrid_linear/test_flashinfer_mla_chunk_metadata.py b/test/registered/attention/unittests/hybrid_linear/test_flashinfer_mla_chunk_metadata.py
index a00851797..e5cebe281 100644
--- a/test/registered/attention/unittests/hybrid_linear/test_flashinfer_mla_chunk_metadata.py
+++ b/test/registered/attention/unittests/hybrid_linear/test_flashinfer_mla_chunk_metadata.py
@@ -45,13 +45,14 @@ class _ChunkKVMLARunner(MockMLAModelRunner):
 
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
-        # The fixture's config is already published; adjust it through the
-        # audited entry point (bare writes raise under the strict guard).
-        self.server_args.override(
-            source="attention-unittest",
+        from sglang.srt.runtime_context import get_context
+
+        self.server_args = self.server_args.derive(
+            "attention-unittest",
             disable_chunked_prefix_cache=False,
             flashinfer_mla_disable_ragged=False,
         )
+        get_context().set_server_args(self.server_args)
 
 
 def _make_case() -> MLAAttentionCase:
diff --git a/test/registered/unit/parser/test_template_manager.py b/test/registered/unit/parser/test_template_manager.py
index b9b257b88..af5e0ea0c 100644
--- a/test/registered/unit/parser/test_template_manager.py
+++ b/test/registered/unit/parser/test_template_manager.py
@@ -13,6 +13,7 @@ from sglang.srt.parser.template_detection import (
     detect_tool_call_parser,
     resolve_auto_parsers,
 )
+from sglang.srt.server_args import ServerArgs
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=2.0, suite="base-a-test-cpu")
@@ -769,21 +770,15 @@ class TestResolveAutoParsers(unittest.TestCase):
 
     qwen3_template = "{% set enable_thinking = enable_thinking if enable_thinking is defined else true %}"
 
-    class _Args(SimpleNamespace):
-        # Write-through override, per the runtime-context testing idiom:
-        # production adjusts parsers through override(source, ...), so the
-        # stand-in needs the method (a bare SimpleNamespace would raise).
-        def override(self, source, **fields):
-            for key, value in fields.items():
-                setattr(self, key, value)
-
     def _make_server_args(
         self, reasoning_parser=None, tool_call_parser=None, chat_template=None
     ):
-        return self._Args(
+        # The dummy model path skips resolution; the tokenizer / HF-config
+        # loads that detection performs are patched per test.
+        return ServerArgs(
+            model_path="dummy",
             reasoning_parser=reasoning_parser,
             tool_call_parser=tool_call_parser,
-            model_path="Qwen/Qwen3-0.6B",
             trust_remote_code=False,
             chat_template=chat_template,
         )
@@ -826,7 +821,11 @@ class TestResolveAutoParsers(unittest.TestCase):
 
     def test_nonexistent_model_disables_both_parsers(self):
         args = self._make_server_args(reasoning_parser="auto", tool_call_parser="auto")
-        args.model_path = "nonexistent/model-does-not-exist-xyz"
+        args = self._make_server_args(
+            reasoning_parser="auto",
+            tool_call_parser="auto",
+        )
+        object.__setattr__(args, "model_path", "nonexistent/model-does-not-exist-xyz")
         with _patch_hf_transformers_utils(
             Mock(side_effect=RuntimeError("tokenizer unavailable")),
             Mock(side_effect=RuntimeError("config unavailable")),
diff --git a/test/registered/unit/test_runtime_context.py b/test/registered/unit/test_runtime_context.py
index d3ac25300..6079f9614 100644
--- a/test/registered/unit/test_runtime_context.py
+++ b/test/registered/unit/test_runtime_context.py
@@ -299,12 +299,13 @@ class TestServerArgsScopedOverride(_IsolatedServerArgs):
 
     def test_installed_config_arms_the_strict_guard(self):
         # The published dummy must behave like a resolved config: bare writes
-        # raise under the strict harness; override() stays the entry point.
+        # raise, and a differing value lives on a derived variant.
         published = get_context().override_server_args(tp_size=2).install()
         with self.assertRaises(AttributeError):
             published.tp_size = 4
-        published.override(source="test", tp_size=4)
-        self.assertEqual(published.tp_size, 4)
+        variant = published.derive("test", tp_size=4)
+        self.assertEqual(variant.tp_size, 4)
+        self.assertEqual(published.tp_size, 2)
 
     def test_restore_resets_the_capture_seed(self):
         # install() seeds flags.capture from the published dummy; restore()
diff --git a/test/registered/unit/test_runtime_context_override.py b/test/registered/unit/test_runtime_context_override.py
index 9f99a89c1..93c344242 100644
--- a/test/registered/unit/test_runtime_context_override.py
+++ b/test/registered/unit/test_runtime_context_override.py
@@ -102,8 +102,8 @@ class TestContextOverride(CustomTestCase):
         self.assertEqual(sa.kv_cache_dtype, raw)
 
     def test_bare_server_args_write_raises_after_resolution(self):
-        # server_args is read-only after resolution regardless of the
-        # SGLANG_STRICT_CONFIG_MUTATION env; write via override instead.
+        # server_args is read-only after resolution: resolved config changes go
+        # to the bags, a per-runner config to a derived variant.
         sa = ServerArgs(model_path="dummy")
         object.__setattr__(sa, "_declarations_materialized", True)
         with self.assertRaises(AttributeError):
@@ -118,8 +118,9 @@ class TestContextOverride(CustomTestCase):
         rc.get_context().override(
             "ModelRunner.configure_kv_cache_dtype", kv_cache_dtype="fp8_e4m3"
         )
-        draft = ServerArgs(model_path="dummy")
-        draft.override(source="draft-build", kv_cache_dtype="bf16")
+        draft = ServerArgs(model_path="dummy").derive(
+            "draft-build", kv_cache_dtype="bf16"
+        )
         with rc.get_context().preserve_config():
             rc.get_context().set_server_args(draft)
             # Inside the scope the draft's bags are live...
diff --git a/test/registered/unit/test_server_args_derive.py b/test/registered/unit/test_server_args_derive.py
new file mode 100644
index 000000000..3ddca37e5
--- /dev/null
+++ b/test/registered/unit/test_server_args_derive.py
@@ -0,0 +1,79 @@
+"""``ServerArgs.derive`` is the only way one config becomes another.
+
+After resolution the instance is the process's read-only startup record and the
+object the config bags were projected from, so it cannot be mutated: a change to
+resolved config goes to the bags (``get_context().override``), and a config that
+differs for one runner or worker — a draft's context length, an encode worker's
+device — is a second object.
+"""
+
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+import unittest
+
+from sglang.srt.runtime_context import get_context, reset_context
+from sglang.srt.server_args import ServerArgs
+from sglang.test.test_utils import CustomTestCase
+
+
+class TestServerArgsDerive(CustomTestCase):
+    def tearDown(self):
+        reset_context()
+
+    def _resolved(self) -> ServerArgs:
+        """A config in the post-resolution state, without a real model."""
+        server_args = ServerArgs(model_path="dummy")
+        object.__setattr__(server_args, "_declarations_materialized", True)
+        return server_args
+
+    def test_the_receiver_is_untouched(self):
+        server_args = self._resolved()
+        variant = server_args.derive("draft_worker.copy", context_length=1024)
+
+        self.assertEqual(variant.context_length, 1024)
+        self.assertIsNot(variant, server_args)
+        self.assertIsNone(server_args.context_length)
+
+    def test_a_published_config_rejects_assignment_but_still_derives(self):
+        override = get_context().override_server_args(tp_size=2)
+        published = override.install()
+        self.addCleanup(override.restore)
+
+        with self.assertRaises(AttributeError):
+            published.tp_size = 4
+
+        self.assertEqual(published.derive("worker", tp_size=4).tp_size, 4)
+        self.assertEqual(published.tp_size, 2)
+
+    def test_the_variant_is_not_published_by_deriving(self):
+        override = get_context().override_server_args(tp_size=2)
+        published = override.install()
+        self.addCleanup(override.restore)
+
+        published.derive("worker", tp_size=4)
+
+        self.assertIs(get_context().server_args, published)
+
+    def test_the_variant_is_frozen_too(self):
+        variant = self._resolved().derive("worker", context_length=1024)
+        with self.assertRaises(AttributeError):
+            variant.context_length = 2048
+
+    def test_provenance_is_recorded(self):
+        variant = self._resolved().derive("draft_worker.copy", watchdog_timeout=1.0)
+        self.assertIn(
+            ("draft_worker.copy", {"watchdog_timeout": 1.0}),
+            getattr(variant, "_runtime_mutations", []),
+        )
+
+    def test_a_resolvable_field_joins_the_declaration_stash(self):
+        """So a re-resolution of the variant keeps the derived value."""
+        variant = self._resolved().derive("draft_worker.build", kv_cache_dtype="bf16")
+        stash = getattr(variant, "_resolved_overrides", [])
+        self.assertIn(("draft_worker.build", {"kv_cache_dtype": "bf16"}), stash)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/test/registered/unit/test_server_args_mutation_ratchet.py b/test/registered/unit/test_server_args_mutation_ratchet.py
index 848c2d78c..4eea1aefe 100644
--- a/test/registered/unit/test_server_args_mutation_ratchet.py
+++ b/test/registered/unit/test_server_args_mutation_ratchet.py
@@ -8,13 +8,13 @@ configuration; the resolution pipeline (``server_args.py`` and
 an exact pin: new mutations must not appear, and removals must lower the
 baseline to lock in the progress.
 
-Every audited runtime adjustment goes through ``ServerArgs.override(source,
-**fields)`` — the single mutation entry point, which records provenance and
-keeps whitelisted fields consistent with the declaration stash. The baseline
-is therefore zero. The registered test harness additionally runs with
-``SGLANG_STRICT_CONFIG_MUTATION=1``, under which a bare assignment after
-resolution raises at runtime; this ratchet catches sites the tests never
-execute.
+There is no post-resolution mutation entry point on the instance any more:
+resolved config changes go to the context bags via
+``get_context().override(source, **fields)``, and a config that differs for one
+runner or worker is a separate object built with ``ServerArgs.derive(source,
+**fields)``. The baseline is therefore zero. ``ServerArgs.__setattr__`` raises
+on a bare assignment after resolution; this ratchet catches the sites the tests
+never execute.
 """
 
 from sglang.test.ci.ci_register import register_cpu_ci
@@ -69,8 +69,9 @@ class TestServerArgsMutationRatchet(CustomTestCase):
                 f"server_args mutations outside the resolution pipeline grew: "
                 f"{count} > baseline {_BASELINE}. Configuration is resolved in "
                 "ServerArgs.__post_init__; declare through the pipeline "
-                "(passes / declare_load_time_override) or go through "
-                "ServerArgs.override(source, ...) instead of assigning fields."
+                "(passes / declare_load_time_override), change resolved config "
+                "with get_context().override(source, ...), or build a variant "
+                "with server_args.derive(source, ...) — do not assign fields."
             )
         if count < _BASELINE:
             self.fail(
diff --git a/test/registered/unit/test_server_args_no_instance_mutation_entry.py b/test/registered/unit/test_server_args_no_instance_mutation_entry.py
new file mode 100644
index 000000000..2656589ad
--- /dev/null
+++ b/test/registered/unit/test_server_args_no_instance_mutation_entry.py
@@ -0,0 +1,85 @@
+"""``ServerArgs`` has no in-place mutation entry, and nothing calls one.
+
+``ServerArgs.override(source, **fields)`` used to mutate a resolved instance;
+after resolution the fields are the record the config bags were projected from,
+so such a write desyncs every namespace reader. The method is gone and the two
+sanctioned replacements are ``get_context().override`` (post-publish, writes the
+bags) and ``ServerArgs.derive`` (a variant for another runner / process). Late
+launcher-stage resolution writes in place through
+``arg_groups.overrides.declare_late_resolution``, which refuses the published
+instance.
+
+The textual half of this guard matters because the resolution pipeline's own file
+is exempt from the mutation ratchet: a ``self.override(...)`` there — exactly
+what the LoRA normalization used — is invisible to it.
+"""
+
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+import re
+import unittest
+from pathlib import Path
+
+import sglang
+from sglang.srt.server_args import ServerArgs
+from sglang.test.test_utils import CustomTestCase
+
+_SGLANG_ROOT = Path(next(iter(sglang.__path__)))
+
+# ``x.override(`` on anything that is a ServerArgs by name, including the
+# pipeline's own ``self.override(`` inside server_args.py.
+_PATTERNS = [
+    re.compile(r"\bself\.override\("),
+    re.compile(r"\bserver_args\.override\("),
+    re.compile(r"\bsa\.override\("),
+    re.compile(r"\bargs\.override\("),
+]
+
+_EXCLUDED = ("multimodal_gen",)
+
+
+class TestNoServerArgsMutationEntry(CustomTestCase):
+    def test_the_method_is_gone(self):
+        self.assertFalse(
+            hasattr(ServerArgs, "override"),
+            "ServerArgs.override is back; post-publish changes belong on the bags "
+            "(get_context().override) and per-runner values on a derive() variant.",
+        )
+
+    def test_nothing_calls_an_instance_override(self):
+        offenders = []
+        for path in sorted(_SGLANG_ROOT.rglob("*.py")):
+            rel = path.relative_to(_SGLANG_ROOT).as_posix()
+            if rel.startswith(_EXCLUDED):
+                continue
+            source = path.read_text()
+            for pattern in _PATTERNS:
+                for match in pattern.finditer(source):
+                    line = source.count("\n", 0, match.start()) + 1
+                    offenders.append(f"{rel}:{line}: {match.group(0)}")
+        if offenders:
+            self.fail(
+                "in-place ServerArgs mutation call-sites:\n"
+                + "\n".join(offenders)
+                + "\n\nUse get_context().override(source, ...) for resolved config, "
+                "server_args.derive(source, ...) for a per-runner variant, or "
+                "declare_late_resolution(...) for pre-publish launcher resolution."
+            )
+
+    def test_late_resolution_refuses_the_published_config(self):
+        from sglang.srt.arg_groups.overrides import declare_late_resolution
+        from sglang.srt.runtime_context import get_context
+
+        override = get_context().override_server_args(tp_size=2)
+        published = override.install()
+        self.addCleanup(override.restore)
+
+        with self.assertRaises(ValueError):
+            declare_late_resolution(published, "test", tp_size=4)
+        self.assertEqual(published.tp_size, 2)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/test/registered/unit/test_server_args_writer_ratchet.py b/test/registered/unit/test_server_args_writer_ratchet.py
deleted file mode 100644
index 6a611dfb6..000000000
--- a/test/registered/unit/test_server_args_writer_ratchet.py
+++ /dev/null
@@ -1,81 +0,0 @@
-"""Ratchet guard: ``ServerArgs.override`` call-sites may only decrease.
-
-``ServerArgs.override(source, **fields)`` mutates a ``ServerArgs`` *instance*
-only — the resolved-config bags on the runtime context never see the write, so
-any consumer reading the namespace accessors (``get_exec()`` / ``get_memory()``
-/ …) desyncs from the writer. The migration end-state removes this primitive
-entirely: post-publish, process-global config changes go through
-``get_context().override(source, **fields)`` (which writes the bags), and
-per-runner resolved values live on the runner object rather than on a
-``ServerArgs`` copy.
-
-Until every call-site is rerouted together with its readers, this exact pin
-keeps the writer surface from growing unwatched: new writers must use
-``get_context().override``, and each rerouted batch lowers the baseline to
-lock in the progress. (The count is textual and includes docstring mentions
-and the test-kit's private-config use — the pin tracks growth, not the exact
-production-writer census.)
-"""
-
-from sglang.test.ci.ci_register import register_cpu_ci
-
-register_cpu_ci(est_time=5, suite="base-a-test-cpu")
-
-import re
-import unittest
-from pathlib import Path
-
-import sglang
-from sglang.test.test_utils import CustomTestCase
-
-_SGLANG_ROOT = Path(next(iter(sglang.__path__)))
-
-# ``server_args.override(`` also matches ``self.server_args.override(``,
-# ``<obj>.server_args.override(``, and the ``draft_server_args`` /
-# ``dp_server_args`` copies; ``args`` / ``sa`` are the aliases a few call-sites
-# bind first.
-_WRITER_PATTERNS = [
-    re.compile(r"server_args\.override\("),
-    re.compile(r"\bargs\.override\("),
-    re.compile(r"\bsa\.override\("),
-]
-
-# The resolution pipeline itself (its declare face forwards through
-# ``override`` by design) and multimodal_gen, whose ServerArgs is a different
-# class outside this contract.
-_EXCLUDED = (
-    "srt/server_args.py",
-    "srt/arg_groups",
-    "multimodal_gen",
-)
-
-_BASELINE = 15
-
-
-class TestServerArgsWriterRatchet(CustomTestCase):
-    def test_server_args_override_call_sites_match_the_baseline(self):
-        count = 0
-        for path in sorted(_SGLANG_ROOT.rglob("*.py")):
-            rel = path.relative_to(_SGLANG_ROOT).as_posix()
-            if rel.startswith(_EXCLUDED):
-                continue
-            source = path.read_text()
-            count += sum(len(p.findall(source)) for p in _WRITER_PATTERNS)
-        if count > _BASELINE:
-            self.fail(
-                f"ServerArgs.override call-sites grew: {count} > baseline "
-                f"{_BASELINE}. Instance writes never reach the resolved-config "
-                "bags, so namespace readers desync from the writer. Post-publish "
-                "process-global changes go through get_context().override(...); "
-                "per-runner resolved values belong on the runner object."
-            )
-        if count < _BASELINE:
-            self.fail(
-                f"ServerArgs.override call-sites shrank: {count} < baseline "
-                f"{_BASELINE}. Lower the baseline in this file to lock in the "
-                "progress."
-            )
-
-
-if __name__ == "__main__":
-    unittest.main()