test: stand up the config tiers two unit tests read from

[ch-wan]

What

Two registered tests fail on main today. Both read process state that only a
served engine establishes, and nothing in the test published it:

• test/registered/chunked_prefill/test_mm_chunked_embedding_unit.py — 3 of 5
  tests fail with AssertionError: attention tensor model parallel group is not initialized. The embedding path reads the attention-TP rank and the config;
  the test set up neither.
• test/registered/unit/hardware_backend/mlx/test_attn_dp_request_capacity.py —
  fails on the MLX runner with ValueError: config namespace 'exec' not published. It handed MlxModelRunnerStub a SimpleNamespace carrying the four
  fields initialize() used to read off the instance; those reads moved to the
  namespace accessors, so the stand-in stopped intercepting.

Failure logs: mm,
mlx.

Fix

Both tests publish a real config instead of faking one, which is what the
runtime-context testing contract asks for — a faked accessor or a stubbed
server_args silently stops intercepting the moment a reader migrates, which is
exactly what happened here. The mm test also overrides the parallel tier
(get_parallel().override(attn_tp_rank=0, ...)) for the duration of each test.
The MLX stub still receives the published instance, so its own instance reads are
unchanged.

While here, one production line: _acknowledge_deferred_cuda_ipc_cache_hits
computed getattr(server_args, "tp_size", parallel.attn_tp_size). Python
evaluates that default eagerly, so every call read the live attention-TP size
even though tp_size is always present on a resolved ServerArgs — a config read
that required a process group for no reason, and the reason the mm test tripped
the assert before reaching anything else. It reads tp_size directly now.

Validation

The two test files: 5 passed + 4 skipped (MLX skips without mlx installed; the
four namespace reads its stub makes were verified to resolve under the new
publish). test/registered/unit/{multimodal,managers} plus the config ratchets:
358 passed. The MLX path itself needs the MLX runner in CI.

Split out of a larger ServerArgs.override burndown stack (#33242) so the
main-side fix can land on its own.

---

CI States

Latest PR Test (Base): 🚫 Run #30770816029
Latest PR Test (Extra): ❌ Run #30770815906

[gemini-code-assist]

Caution

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

[ch-wan]

Renamed the fixture to publish_config_and_parallel_state and said in its docstring that it is autouse — applied to every test in the module and named by none of them, which makes it look unused on a grep.

It is load-bearing: flipping autouse=True to False puts the file straight back to 3 failed / 2 passed, and restoring it returns 5 passed.