feat: add ImageClassificationExecutor for vision fine-tuning

[J1shen]

Purpose

FlowMesh's training executors (SFT/LoRA/DPO/PPO) are all LLM fine-tuning. This PR adds a first-class image-classification training task type so vision models can be fine-tuned through the same workflow/submit path, instead of users wrapping their own trainer in an ssh container job.

Changes

• src/shared/tasks/task_type.py — add IMAGE_CLASSIFICATION_TRAINING = "image_classification_training".
• src/shared/tasks/specs/training.py — add ImageClassificationTrainingSpec{Strict,Template}; the strict spec requires model.source.identifier (validated, no silent default).
• src/shared/tasks/specs/__init__.py, src/shared/tasks/envelope.py — export the spec and wire it into the taskType discriminator union (strict + template).
• sdk/src/flowmesh/models/common.py — mirror the new enum member (kept in lockstep with the server enum, enforced by tests/sdk/test_schema_compat.py).
• src/worker/executors/image_classification_executor.py — new ImageClassificationExecutor: AutoImageProcessor + AutoModelForImageClassification + HF Trainer. Loads the dataset via datasets, infers num_labels/id2label, trains, evaluates, saves final_model/. Honors training.optimizer (→ TrainingArguments.optim) and training.augmentation (h-flip + 4px-pad random crop on the train split), and returns the per-step train_losses. Reuses the existing checkpoint/artifact helpers (determine_resume_path, archive_model_dir, maybe_upload_artifacts, TrainingMixin._cleanup_local_artifacts).
• src/worker/executors/__init__.py — register image_classification_training in EXECUTOR_REGISTRY / EXECUTOR_CLASS_NAMES / __all__.
• src/worker/main.py — instantiate the new executor on worker startup: add image_classification_training to the executor build list and to _EXECUTORS_TO_WRAP (training executors run in an MPExecutor subprocess for GPU cleanup). Without this the worker registered the type but fell back to the transformers executor and rejected the spec.
• src/server/task/models.py — add image_classification_training to TRAINING_TASK_TYPES so the server classifies/routes it as a training task (per review feedback).
• docs/EXECUTORS.md — add the taskType row; relabel the existing training row as LLM fine-tuning.
• examples/templates/image_classification_vit.yaml — runnable ViT-on-CIFAR-10 demo.
• tests/worker/test_executor_registry.py, tests/shared/test_image_classification_spec.py — registry key, _EXECUTORS_TO_WRAP guard, envelope-resolution and required-model coverage.

Dependencies

• pyproject.toml, uv.lock, src/server/requirements.txt, src/worker/requirements/requirements.txt — raise the aiohttp floor >=3.13.4 → >=3.14.0 (re-locked to 3.14.1) and regenerate the worker requirements. This is not a feature change: the strict server pip-audit job started failing on aiohttp 3.13.5 for GHSA-jg22-mg44-37j8 and GHSA-hg6j-4rv6-33pg (both fixed in 3.14.0), which blocked this branch's CI. Bundled here only to keep the pipeline green; happy to split it into its own PR if preferred.

Design

The executor mirrors SFTExecutor's structure (spec → dataset prep → Trainer → save → result) so it slots into the existing training conventions, and deliberately adds no new feature dependencies — it reuses the runtime-inference group (transformers/torch/datasets/Pillow). It is single-GPU for now: allow_multi_gpu is downgraded with a warning (same as LoRASFTExecutor); multi-GPU via a torchrun/DeepSpeed dist-entry mirroring sft_dist_entry.py is a follow-up. The model_arch-free, HF-identifier-driven design means any AutoModelForImageClassification checkpoint (ViT/ResNet/ConvNeXt/…) works without code changes.

Test Plan

• uv run pre-commit run --all-files
• uv run pytest tests/ --ignore=tests/worker/test_mp_executor_cleanup_gpu.py
• uvx bandit==1.9.4 -c pyproject.toml -r src/
• End-to-end through a real stack (no Docker worker image): brought up a local server + an in-venv GPU worker and drove 4 optimizer-selected variants as image_classification_training workflows on Fashion-MNIST (H100). Confirmed server validation → best_fit dispatch → executor run → result via the control plane; also ran the executor in-process for an AdamW-vs-SGD optimizer contrast.

Test Result

• pre-commit: all hooks pass (gitleaks/isort/black/ruff/codespell/mypy).
• pytest: 990 passed.
• bandit: no issues.
• pip-audit: passes with the aiohttp floor raised.
• Stack run: all 4 variants dispatched to the worker and trained through FlowMesh (eval_accuracy/train_losses returned in the result payload); the in-process AdamW run converged (loss 3.16→1.16) while SGD at the same budget stayed flat (~3.0–3.4), confirming training.optimizer is honored.

---

Pre-submission Checklist

• I have read the contribution guidelines.
• I have run pre-commit run --all-files and fixed any issues.
• I have added or updated tests covering my changes (if applicable).
• I have verified that uv run pytest tests/ passes locally.
• If I changed shared schemas or proto definitions, I have checked downstream compatibility across Server and Worker.
• If I changed the SDK or CLI, I have verified the affected packages work (uv sync --all-packages --group ci --frozen).
• If this is a breaking change, I have prefixed the PR title with [BREAKING] and described migration steps above.
• I have updated documentation or config examples if user-facing behavior changed.

[timzsu]

FlowMesh is intended to be a "service fabric for composable LLM workflows". Could you include some justification on whether you introduce the ViT training path instead of Omni LLM training (e.g. FlowGRPO training for Qwen-Image, etc.)?

In addition, the PR description looks like LLM generated and do not follow the template. Please fix :)

[timzsu]

It seems that there are a few places not following the code style guidelines. Could you take a look at the doc? I have left comments in places that I spotted.

By the way, in ./src/server/task/models.py, please extend TRAINING_TASK_TYPES to include the new executor class.

[timzsu]

LGTM. @kaiitunnz Can you take a look at this PR?

[kaiitunnz]

Several comments. PTAL.

[kaiitunnz]

LGTM.