[hparams] refactor: split training_args.py into per-algorithm package

[Jayce-Ping]

Summary

• Convert the 989-line monolithic training_args.py into a package (training_args/) with one file per algorithm, improving isolation for parallel development and reducing merge conflicts as the algorithm count grows
• Zero breaking changes: all existing import paths (from .training_args import X) resolve identically via the package __init__.py
• Update ff-new-algorithm skill and sample_lifecycle.md docs to reflect the new file paths

Package Structure

src/flow_factory/hparams/training_args/
├── __init__.py   # re-exports (preserves all existing import paths)
├── _base.py      # EvaluationArguments, TrainingArguments, shared utilities
├── _registry.py  # _TRAINING_ARGS_REGISTRY + get_training_args_class()
├── grpo.py       # GRPOTrainingArguments
├── nft.py        # NFTTrainingArguments
├── awm.py        # AWMTrainingArguments
├── dpo.py        # DPOTrainingArguments
├── dgpo.py       # DGPOTrainingArguments (extends GRPO)
└── crd.py        # CRDTrainingArguments

Test plan

• python -c "from flow_factory.hparams import GRPOTrainingArguments, get_training_args_class; print('OK')"
• python -c "from flow_factory.hparams import get_training_args_class; cls = get_training_args_class('grpo'); print(cls.__name__)"
• Verify existing YAML configs load correctly via Arguments.load_from_yaml()
• grep -rn "from.*training_args" src/ shows no broken imports
• Run existing test suite

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR refactors flow_factory.hparams.training_args from a single monolithic module into a package with per-algorithm modules, while preserving the existing public import surface via training_args/__init__.py.

Changes:

• Split training_args.py into training_args/ with _base.py (shared args/utilities), _registry.py (lookup/registry), and one file per algorithm.
• Preserve existing import paths by re-exporting all public symbols from training_args/__init__.py.
• Update agent skill and lifecycle docs to reflect the new file locations.

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/flow_factory/hparams/training_args/_base.py	Introduces shared EvaluationArguments/TrainingArguments and shared standardization utilities.
src/flow_factory/hparams/training_args/_registry.py	Adds centralized registry and dynamic lookup for algorithm-specific TrainingArguments subclasses.
src/flow_factory/hparams/training_args/__init__.py	Re-exports classes/functions to keep existing imports working.
src/flow_factory/hparams/training_args/grpo.py	Adds GRPOTrainingArguments in its own module.
src/flow_factory/hparams/training_args/nft.py	Adds NFTTrainingArguments in its own module.
src/flow_factory/hparams/training_args/awm.py	Adds AWMTrainingArguments in its own module.
src/flow_factory/hparams/training_args/dpo.py	Adds DPOTrainingArguments in its own module.
src/flow_factory/hparams/training_args/dgpo.py	Adds DGPOTrainingArguments (extends GRPO) in its own module.
src/flow_factory/hparams/training_args/crd.py	Adds CRDTrainingArguments in its own module.
src/flow_factory/hparams/training_args.py	Removes the old monolithic module after splitting.
.agents/skills/ff-new-algorithm/SKILL.md	Updates the “new algorithm” agent skill to match the new package structure and registration steps.
.agents/knowledge/topics/sample_lifecycle.md	Updates docs to reference TrainingArguments in training_args/_base.py.

Comments suppressed due to low confidence (1)

src/flow_factory/hparams/training_args/_base.py:278

• Same issue as EvaluationArguments: the width override warning is emitted but self.resolution is never updated, so width= cannot override the resolution and self.width will not reflect the user-provided value.

        if self.width is not None and self.resolution[1] != self.width:
                logger.warning(
                    f"Both `resolution={self.resolution}` and `width={self.width}` are set. "
                    f"Using width to override: ({self.resolution[0]}, {self.width})."
                )

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 12 out of 12 changed files in this pull request and generated 2 comments.

Comments suppressed due to low confidence (1)

src/flow_factory/hparams/training_args/_base.py:276

• The height/width override block in TrainingArguments.__post_init__ is over-indented (extra indentation level), which is inconsistent with the surrounding code and will be reformatted by Black. Adjust indentation to a single level under each if.

        if self.height is not None and self.resolution[0] != self.height:
                logger.warning(
                    f"Both `resolution={self.resolution}` and `height={self.height}` are set. "
                    f"Using height to override: ({self.height}, {self.resolution[1]})."
                )
                self.resolution = (self.height, self.resolution[1])
        if self.width is not None and self.resolution[1] != self.width:
                logger.warning(

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.