When a MaxText model is served through vLLM, RoutedMoE.fused_moe_matmul calls
tpu-inference's fused_moe_func with only the required arguments. tpu-inference's
own serving path passes several environment-backed knobs as well, so the two
paths run the same kernel with different configurations, and env vars that
operators set on the vLLM command line are silently ignored on the MaxText path.

Forward the four environment-backed knobs: ENABLE_RS_KERNEL,
USE_GMM_FUSED_RS_KERNEL, ONEHOT_MOE_PERMUTE_THRESHOLD and VLLM_MOE_CHUNK_SIZE.
ONEHOT_MOE_PERMUTE_THRESHOLD matters beyond tuning: it selects the one-hot
permute path over the SparseCore ragged_gather_reduce kernel, which is what lets
expert parallelism run at all on TPU generations whose SparseCore has fewer SIMD
lanes than that kernel needs. On v5p, expert_parallelism > 1 previously failed
with "num_row_partitions=16 must be <= num_simd_lanes=8".

Also make two silent sharding surprises visible in the vLLM adapter.

Raise the MoE padding message from absl INFO to WARNING and include its cost.
vLLM's logging configuration does not surface absl INFO records, so the padding
was invisible: serving Qwen3.5-35B-A3B at moe_mlp_tp_size=4 pads
moe_intermediate_size 512 -> 1024 and doubles the MoE weights (measured
33.92 GiB/device instead of 16.95 GiB/device), and at moe_mlp_tp_size=8 it
quadruples them, with no indication in the logs.

Warn when --enable-expert-parallel is passed but the mesh has no expert shards.
The native tpu-inference model paths derive use_ep from
parallel_config.enable_expert_parallel, while the mesh takes expert parallelism
only from additional_config's sharding_strategy and MaxText derives use_ep from
the mesh. The same command line therefore runs expert-parallel MoE natively and
tensor-parallel MoE under MaxText, which is the configuration that triggers the
padding above. The warning points at expert_parallelism as the fix.

Bump the post-training pins to pick these up. USE_GMM_FUSED_RS_KERNEL does not
exist in the previously pinned tpu-inference (7ecc401e, Jul 27), and neither
does the use_gmm_fused_rs_kernel argument of fused_moe_func, so the change above
needs a newer commit to import at all:

  tpu-inference  7ecc401e -> b67ae5f8 (main)
  vllm           0ba2aa35 -> d626108b (tpu-inference .buildkite/vllm_lkg.version)

Both the old and new tpu-inference pin jax==0.11.0, jaxlib==0.11.0 and
libtpu==0.0.44, which already match tpu_post_train_overrides.txt, so the jax
version in CI is unchanged.

The post-training lock is regenerated with seed-env using the same JAX seed
commit recorded for it in docs (52d5cb38). Regeneration needs one fix to run:
base requirements ask for an unpinned llguidance, which now resolves to 1.8.0,
while vLLM requires >=1.7.0,<1.8.0 in both the old and new commits. Bound it to
<1.8.0. The functionally required outcome of the regeneration is
tokamax>=0.0.13 (up from 0.0.12, where the GMM_v2 kernel lives) and
huggingface-hub>=1.29.0; the lock is installed with --resolution=lowest and the
GitHub deps with --no-deps, so these floors are exactly what CI installs.