Stabilize Qwen3 MoE LoRA mixed-scale test

[taivu1998]

Summary

• Stabilizes the flaky Qwen3 MoE LoRA comparison reported in Flaky test_qwen3_moe_layer_lora, replace np.allclose? #1604 by making this test input deterministic.
• Replaces the fixed absolute tolerance at the flaky assertion with a local mixed-scale assertion helper that preserves the existing relative tolerance while scaling absolute tolerance with the output dynamic range.
• Adds explicit non-finite checks and a richer failure message so future failures report the adapter, sample, dynamic range, and effective tolerance.

Root Cause

Issue #1604 reports that tests/tx/models/test_qwen3.py::test_qwen3_moe_layer_lora compares tensors whose values span from small numbers to roughly 1e4. A fixed atol=1e-3 can make numerically equivalent JAX fused-LoRA and merged-weight paths fail on small-magnitude entries when the layer output has a much larger dynamic range.

Changes

• Seed only this test input with a local torch.Generator so the flaky random input is reproducible without changing global RNG state.
• Add assert_allclose_mixed_scale for this activation-level comparison.
• Keep the assertion local to the Qwen3 MoE LoRA test and avoid production MoE/LoRA behavior changes.

Validation

• git diff --check -- tests/tx/models/test_qwen3.py
• uv run --with ruff ruff check tests/tx/models/test_qwen3.py
• .venv/bin/python -m pytest -q tests/tx/models/test_qwen3.py::test_qwen3_moe_layer_lora - 3 passed
• OBJC_DISABLE_INITIALIZE_FORK_SAFETY=YES .venv/bin/python -m pytest -q --forked tests/tx/models/test_qwen3.py::test_qwen3_moe_layer_lora - 3 passed

Fixes #1604

[gemini-code-assist]

Code Review

This pull request introduces a new utility function, assert_allclose_mixed_scale, to the Qwen3 model tests for more precise numerical comparisons using a combination of base and scale-dependent absolute tolerances. Additionally, it updates test_qwen3_moe_layer_lora to use a seeded random generator for input data, ensuring test reproducibility. I have no feedback to provide.