sync: merge upstream/master (2026-04-25)

[RyderFreeman4Logos]

Summary

• Synced with upstream/master (11 commits merged)
• Upstream commits: reproducibility system, configurable residual processing, gemma 4 support, --help lazy import, UnboundLocalError fix, response prefix config, max_memory fix, dependency bumps
• 7 conflict files resolved, new system.py accepted
• Fork features preserved: LLM judge, caching, geometric median, thinking profiles, sequence KL, pipelined optimization

Quality Gates

• ruff format
• ruff check
• ty check
• uv build
• lefthook (all 5 hooks)

Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces a comprehensive reproducibility framework for the heretic tool, enabling the capture and upload of system metadata, environment details, and exact configuration settings to Hugging Face. It also implements VRAM optimizations through CPU offloading of intermediate tensors and adds support for sharded model exports. The review feedback identifies a critical NameError caused by an undefined function call, a logic error in the early help-handling mechanism that may trigger validation failures, and several violations of the project's type-annotation requirements for function signatures.

[gemini-code-assist]

The Settings() call here is likely intended to trigger a help message and exit when -h or --help is passed. However, pydantic-settings does not automatically print help and exit upon instantiation unless specifically configured (e.g., via cli_parse_args=True in model_config). If not so configured, this call will likely raise a ValidationError if required fields like model are missing from environment variables or config files. Furthermore, if it doesn't exit, the script will proceed to import heavyweight dependencies on line 20, defeating the purpose of this "lazy help" optimization.

[gemini-code-assist]

The function count_completed_trials is not defined in this module nor imported from any of the internal modules (.config, .evaluator, .model, .system, .utils). This will cause a NameError at runtime when attempting to upload reproducibility information.

[gemini-code-assist]

The function signature is missing the return type annotation. Per the repository style guide (rule 5), all function and method signatures must be fully type-annotated.

Suggested change

	def empty_cache():
	def empty_cache() -> None:

References

1. Function and method signatures must be fully type-annotated, including the return type (if any). ^(link)

[gemini-code-assist]

The function signature is missing the return type annotation. Per the repository style guide (rule 5), all function and method signatures must be fully type-annotated.

Suggested change

	def set_seed(seed: int):
	def set_seed(seed: int) -> None:

References

1. Function and method signatures must be fully type-annotated, including the return type (if any). ^(link)

[gemini-code-assist]

The function signature is missing the return type annotation. Per the repository style guide (rule 5), all function and method signatures must be fully type-annotated.

def create_reproduce_folder(\n    path: Path,\n    settings: Settings,\n    checkpoint_path: str | Path,\n    trial: Trial,\n    uploaded_model_hashes: dict[str, str],\n    include_system_information: bool,\n) -> None:

References

1. Function and method signatures must be fully type-annotated, including the return type (if any). ^(link)

[gemini-code-assist]

The function signature is missing the return type annotation. Per the repository style guide (rule 5), all function and method signatures must be fully type-annotated.

def upload_reproduce_folder(\n    repo_id: str,\n    settings: Settings,\n    token: str,\n    checkpoint_path: str | Path,\n    trial: Trial,\n    include_system_information: bool,\n) -> None:

References

1. Function and method signatures must be fully type-annotated, including the return type (if any). ^(link)

[RyderFreeman4Logos]

PR Bot Audit Trail — Bot Findings Classification

Local CSA review: PASS (gemini-cli, session 01KQ37YY6CQKRMHNC7DZ22CW9F)
Cloud bot: gemini-code-assist — 6 findings (2 HIGH, 4 MEDIUM)

Classification: All 6 findings are FALSE POSITIVES

HIGH #1 — Settings() for --help (main.py:18)
Classification: FALSE_POSITIVE
Rationale: This is upstream's lazy-help optimization pattern. _is_help_invocation() guards it. pydantic-settings with cli_parse_args handles --help correctly at instantiation. Upstream PR p-e-w#293 specifically added and tested this.

HIGH #2 — count_completed_trials undefined (main.py:1651)
Classification: FALSE_POSITIVE
Rationale: count_completed_trials() is defined as a closure inside run() at line 1102: def count_completed_trials() -> int:. The bot failed to detect the nested function definition. All 6 call sites (lines 1106, 1145, 1177, 1434, 1651) are within run()'s scope.

MEDIUM #3-6 — Missing return type annotations (system.py:26, utils.py:345, utils.py:612, utils.py:685)
Classification: FALSE_POSITIVE
Rationale: These are from upstream code accepted as-is. The project uses ty (not mypy) as its type checker, and ty check --error-on-warning passes cleanly. The ty checker does not require return type annotations on all function signatures. No project style guide mandates this — the bot inferred a non-existent "rule 5".

Quality Gates

• ruff format — passed
• ruff check — passed
• ty check --error-on-warning — passed
• uv build — passed
• lefthook pre-commit — all 5 hooks passed
• Local CSA review — PASS (no findings)
• Cloud bot — all findings classified as false positive