fix: normalize output-axis DoRA with adapted weights

[orangemagic123]

Summary

• normalize output-axis DoRA weights using the adapted weight W + ΔW
• keep the existing input-axis behavior unchanged
• add regression coverage for linear and Conv2d-shaped weights

Problem

weight_decompose() currently builds the adapted weight first:

weight_calc = weight + lora_diff

but the output-axis branch calculates its norm from the original base weight:

weight_norm = norm(weight)

This applies:

m * (W + ΔW) / ||W||

instead of the DoRA decomposition:

m * (W + ΔW) / ||W + ΔW||

The input-axis branch already normalizes weight_calc, so this change also makes both branches consistent. The behavior now matches the current LyCORIS LoKr/DoRA implementation.

Tests

A focused CPU regression harness matching the added assertions passed both cases:

linear: passed
conv2d: passed

The added repository test can be run with:

pytest tests-unit/comfy_test/test_weight_adapter.py

The full ComfyUI test suite was not run in the connector environment; GitHub CI is expected to provide repository-level validation.

Related work

This is a separate semantic-correctness fix from the dora_scale broadcasting issue tracked in #12938 and PR #13011. The broadcasting fix may touch the same function, but the adapted-weight normalization remains required after that change.

[github-actions]

✅ All contributors have signed the CLA. Thank you! This PR is ready to be merged.
_{Posted by the CLA Assistant Lite bot.}

[orangemagic123]

I have read and agree to the Contributor License Agreement

[orangemagic123]

recheck