[Bugfix][SM70] Isolate MXFP4 direct-order offsets

[yangzhuxinyzx]

Purpose

Repair the new quality bug reported on #308 without replaying its stale optimization stack. Current main already contains the accepted direct-order optimization from #314, but B1 graph capture can observe offsets overwritten by M2--M8 warmup or verifier compaction.

Source contract

• Base: 9bbc554e1d9f63a628d1bc66ecb45ee4632267a2
• Head: f096a85bf43f3d00df3b25a34299c748ad747b2a
• Original report/fix commit: 052312c66f89bb144cf68f5776d83552c7b9e3d1

Fix

• B1 direct-order decode now uses the immutable slot_expert_offsets buffer ([0, 1, ..., 6]).
• It no longer reuses compact_expert_offsets, which M2--M8 compaction mutates before a B1 graph may be captured.
• CUDA kernels, route IDs/order, FP32 weighted-reduction order, environment defaults, and rollback switches are unchanged.
• A CPU regression poisons the compact workspace with representative M8 offsets and proves direct-order still selects the immutable one-row slots.

The first pre-fix matched endpoint reached 69.949 token/s but produced incoherent output. It is recorded as rejected quality evidence, not a speed baseline. This repair preserves the already measured optimization while removing the identified cross-shape state leak.

Test result

• pytest -q tests/quantization/test_sm70_mxfp4_moe.py with CUDA hidden: 25 passed, 4 skipped.
• pre-commit run --from-ref onecat/main --to-ref HEAD: passed, including Ruff, Markdown, mypy, SPDX, configuration, and forbidden-API checks.
• No GPU or end-to-end rerun: this change only selects an existing immutable buffer before the unchanged CUDA calls, and all V100s remain owned by the active service.

Supersedes the post-#314 bugfix portion of #308 after merge.