config: six more runtime readers ask the bags

[ch-wan]

The series

Stacked, each PR based on the one above it. Read them in order; every boundary is
self-sufficient and green on its own.

1. #36896 — the resolution pipeline's dispatcher leaves the record
2. #36972 — the callbacks into the record go to zero
3. #36973 — six more runtime readers ask the bags ← you are here
4. #36974 — the dead record parameters go
5. #36975 — the lazy imports that buy nothing become eager

• #36925 — CI vehicle — runs the whole series against main, not for merge

Motivation

The previous PR took the resolution package off the record. This one is about the runtime:
code that runs after publish, holds a ServerArgs, and reads raw fields off it.

A census of srt/ finds 295 such reads in 53 files, every field of them in a published namespace.
Not all are convertible, and the criterion is not the field — it is the process. The launcher trio
(http_server, engine, data_parallel_controller) reads the record before its process
publishes; those 117 reads have to stay, and this PR does not touch them.

expert_distribution.py is the clearest case of the rest. ExpertDistributionRecorder was handed
a ServerArgs and threaded it through six constructors — recorder, gatherer, accumulator and
their subclasses — to reach four reads. Two of the classes stored it; one subclass took it and
never touched it. The file already asked get_exec().moe for the recorder mode, so the record was
travelling alongside a bag that already answered.

Modifications

The EPLB recorder. The four reads go to the bag and the parameter goes with them: five call
sites in model_runner.py, six signatures, two stored attributes.

eplb_min_rebalancing_utilization_threshold is an exec.moe leaf. The three balancedness
predicates were members of the record — should_log_expert_balancedness_to_server_log and its
Prometheus sibling, read from four places across three files, each through a record someone was
holding for that purpose. They become published accessors in runtime_context, the way
uses_mla_backend and cutedsl_moe_max_num_tokens already pair with their pre-publish
equivalents. Both members are gone from ServerArgs.

resolve_collector_class was the one thing keeping the record in the chain: it took a ServerArgs
to read stat_loggers, an observability leaf, and three of its six callers were reaching for
get_server_args() to pass it straight back. It asks the bag now, and server_args=None — which
meant "no config, use the default" — becomes "the namespace is not published", the same case in
the context's own terms.

Six more runtime readers, each already reading its other configuration from a bag:

file	reads	bag
hybrid_pool_assembler	6 × hicache_host_memory_mode	memory
scheduler	3 × hicache_host_memory_mode	memory
kt_ep_wrapper	7 × six kt_* leaves	exec.moe
expert_backup_client	nnodes, node_rank	parallel
dspark_planner	three speculative_dspark_*	spec
prefill_delayer	three prefill-delay leaves	schedule

Three of them were carrying the record only for those reads, so it goes: PrefillDelayer and
ExpertBackupClient lose the constructor parameter and the stored attribute, and
build_sps_cost_table — whose body was already entirely get_spec() and get_schedule() — loses
its keyword-only server_args, which takes it out of DsparkPlanner too.

Removing a constructor parameter reaches further than the file that declares it. Three call sites
outside srt/ had to follow: the DSpark worker constructs DSparkVerifyPlanner (a P1 — every
DSPARK startup would have raised TypeError), and the registered prefill-delayer and SPS-table
tests both passed the keyword. The delayer test now publishes its per-case schedule leaves through
override_server_args instead of handing in a stand-in.

Dropping the argument leaves the local that fed it: base_prefix_cache.init_metrics_collector
still bound server_args = get_server_args() for a call that no longer takes one. Its two sibling
sites had already lost both.

The accessor docstrings say what a caller needs and no more: reports_expert_balancedness names
overrides.should_report_expert_balancedness as its pre-publish equivalent, the way
cutedsl_moe_max_num_tokens already did, and drops the sentence about why the accessors exist.

TestResolveCollectorClass drove resolve_collector_class with a _StubArgs stand-in. A stand-in
proves nothing about a function that reads a bag, so the cases force the leaf through
override_server_args and the stand-in is deleted.

Accuracy Tests

The 62-shape resolution probe is byte-identical to the base commit. The affected suites
(observability, mem_cache, managers, spec, server_args) report the same failure set as
main, file for file — 3038 passed on both sides, with the same three pre-existing failures.

Speed Tests and Profiling

None. A bag read is a dataclass attribute read.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

Small and readable in full. The judgement call worth checking is which files were left alone: the
launcher trio publishes, so its reads are pre-publish by necessity, and roughly 150 reads across
thirty smaller files still need that per-file timing check before they can move.

---

CI States

Latest PR Test (Base): 🚫 Run #33249843844
Latest PR Test (Extra): 🚫 Run #33249843729
Latest PR Test (AMD ROCm 7.2): ❌ Run #33249843794