config: retire the last process-global config field reads

[ch-wan]

Part 5 of five, all based on main and meant to merge in order.
Applies on top of parts 1–4. Until they land, this PR's diff includes them.

Replaces #33244, which was closed unmerged: GitHub treated the previous
chained-base series as a stack, which blocks both base retargeting and
every merge path except the async endpoint. The review discussion and the
triage of each round of comments is on #33244; the code here is identical to
that PR's final revision.

What

get_server_args().<field> reads the startup record of one process. Twenty-six of
those reads were left in srt; this removes every one that has somewhere better to
go, and pins the rest.

Nine reads move to the namespace accessors — the value they want is the resolved
one, including post-publish overrides: attention_backend ×5 →
get_exec().kernel, skip_tokenizer_init ×2 → get_serving(), the draft-aware
load_format → get_model(), and a PP size → get_parallel() (live topology, which
is what the DSA indexer wants).

Two reads move to the instance instead of a namespace. The multimodal processor
picked its fast-image-processor device from the process-global config, but
base_gpu_id differs per worker — the encode-server DP workers each specialise their
own copy — so no process-global value can stand in for it, and several Engines can
share a tokenizer process. The decision moves into
BaseMultimodalProcessor._fast_image_processor_device, reading the instance the
processor was constructed with, like the ten other config reads in that class. Branch
order, the NPU preprocess patches and the case that leaves device unset are
unchanged.

What stays is the derived API: mamba_cache_chunk_size (@property),
get_model_config(), enable_mamba_extra_buffer{,_lazy}() — computed from several
fields plus the HF config, so they are not namespace leaves and ServerArgs is their
only home. Two config-intent reads of live-shadowed sizes also stay: get_parallel()
shadows tp/pp/dcp/attn_cp/moe_dp_size with the live topology, and both sites run
where no process group exists (cuda_ipc_transport_utils even guards for "not
published yet").

A new AST ratchet pins field reads at zero and exempts the above by name, with the
reasoning in its docstring.

Test changes

test_dllm_fdfo_kv_reuse rebound allocation.get_server_args to a
SimpleNamespace. That kind of stand-in stops intercepting the moment a reader
migrates — which is exactly what happened here — so it publishes a real config now.
New test_processor_device_selection.py pins the per-instance device resolution
against a conflicting published config.

Validation

Per-area unit suites (mem_cache, multimodal, batch_overlap,
layers/attention, managers) and all ratchets pass. Full registered CPU battery
(16 partitions) against main: seven branch-only failures, each re-run serially and
green — two were port collisions from the parallel run itself, the rest are the
box's standing environment failures. No new failures.

Not covered locally: the flipped reads sit in model and attention paths that need a
GPU to execute; the flips are value-preserving (the bags are projected from the same
published config), but the speculative and model CI suites are the real check.

---

CI States

Latest PR Test (Base): 🚫 Run #30784352926
Latest PR Test (Extra): 🚫 Run #30784352743

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.