security: unblock Torch and Setuptools training runtime alerts

[w0rldx]

Problem

Dependabot alerts #14 (torch, low) and #15 (setuptools, medium) in tools/training/uv.lock cannot currently be fixed independently without breaking the shipped training runtime:

• current torch 2.11.0+cu128 declares setuptools<82, while alert build(deps-dev): bump the frontend-openapi-codegen group in /XE-Local-AI-Engine.Client.React with 2 updates #15 requires setuptools 83.0.0+
• alert build(deps): bump the frontend-runtime group in /XE-Local-AI-Engine.Client.React with 5 updates #14 requires Torch 2.13.0+
• released Unsloth still requires torch<2.12, and released Unsloth Zoo requires torch<2.13
• the compatible PyTorch pair is Torch 2.13.0 + torchvision 0.28.0; official Python 3.13 Linux x86-64 wheels are available on CUDA 12.9, not the current CUDA 12.8 lane

Do not use --no-deps, edit wheel metadata, force setuptools 83 beside Torch 2.11, or ship a custom PyTorch backport.

Upstream unblock conditions

Proceed only after released PyPI artifacts establish all of the following:

1. Unsloth admits Torch 2.13.
2. Unsloth Zoo admits Torch 2.13.
3. The released pair still admits Python 3.13 and the repository's Transformers/TRL/PEFT/xformers closure.
4. Official CUDA 12.9 CPython 3.13 x86-64 Torch 2.13 / torchvision 0.28 wheels remain available.
5. Upstream or local evidence demonstrates Unsloth CUDA training on Torch 2.13; metadata widening alone is insufficient.

Upstream tracking:

• [Bug] Published constraints block torch 2.13 security remediation (GHSA-rrmf-rvhw-rf47) unslothai/unsloth#8926
• deps: widen CUDA/CPU torch ceiling to admit 2.13.x unslothai/unsloth#9003
• Build(deps): bump the python group across 1 directory with 6 updates unslothai/unsloth-zoo#929

Future implementation lane

In one dedicated PR, update together:

• torch==2.13.0
• torchvision==0.28.0
• first exact released compatible Unsloth and Unsloth Zoo versions
• PyTorch source/index from cu128 to cu129
• setuptools==83.0.0 or the minimal later resolver-compatible safe release

Regenerate tools/training/uv.lock, verify all Torch artifacts come from the official cu129 index, then validate imports/versions, CUDA 12.9/RTX 5090 device detection, bitsandbytes and xformers native operations, a real deterministic QLoRA job with measured GPU use, checkpoint/export, and staged llama.cpp GGUF conversion/promotion. Finish with repository Python, targeted training, Release build/test, CI, and Dependabot rescans.

Current disposition

Leave both alerts open as upstream trackers. They are not false positives. If policy later requires dashboard suppression before upstream support exists, use only a documented temporary tolerable_risk dismissal linked to this issue.