Use the configured FFN drop path for ViT residuals

[zjn20030811]

Summary

• Route the FFN residual through the block's configured drop_path2 in the moderate stochastic-depth branch.
• Preserve the historical drop_path1 fallback for direct callers that omit drop_path2.
• Add focused regression coverage for distinct path selection and the compatibility fallback.

User-Visible Behavior

Transformer training now applies the independently configured stochastic-depth module to the FFN residual. Evaluation behavior and callers that omit the optional second path are unchanged.

Affected Pipeline / Benchmark

• Pipeline(s): N/A (shared ViT transformer primitive)
• Benchmark(s): N/A
• Runtime profile(s): N/A
• Public entrypoint(s): worldfoundry.core.nn.vit_block.apply_prenorm_transformer_residuals

Change Type

• Bug fix
• Model integration
• Benchmark integration
• Pipeline/runtime change
• Documentation only
• Test/QA tooling

Asset / API / GPU Requirements

• Downloads or large assets: None
• API keys or quota: None
• GPU / simulator requirements: None; validation is CPU-only
• Official repository or checkpoint assumptions: None

Commands Run

Command	Result	Notes
python -m pytest -q test/eval_core/test_vit_block.py	Pass	2 passed
python -m py_compile worldfoundry/core/nn/vit_block.py test/eval_core/test_vit_block.py	Pass	Both changed Python files compile
git diff --check	Pass	No whitespace errors

Checkpoint / API Key Needs

• Checkpoints / weights: None
• Required environment variables: None
• API providers: None
• Local cache assumptions: None

Sample Artifact Evidence

• Artifact path or link: None
• What it demonstrates: Deterministic CPU regression output and invocation counts
• Known limitations: Full model training and GPU kernels were not run

Validation Matrix Status

Area	Status	Evidence / Command
Unit or focused regression	Pass	python -m pytest -q test/eval_core/test_vit_block.py (2 passed)
Real inference validation	N/A	Shared residual helper only; no checkpoint path changed
Streaming or multi-turn path	N/A	Not applicable
Benchmark runner / metric path	N/A	Not applicable
Docs build or link check	N/A	No documentation changed
GPU/API-key dependent path	N/A	CPU-only change

Leaderboard Validity Impact

• Readiness status changed: No
• leaderboard_valid impact: None
• Scorecard/preflight evidence: Not applicable
• Demo or contract-only limitations: None

Compatibility And Risk

• Backward compatibility impact: drop_path2=None continues to use drop_path1; evaluation and high-ratio stochastic-depth branches are unchanged.
• Expected resource cost: None beyond the existing FFN drop-path operation.
• Failure modes or rollout concerns: The fix changes training-time residual masking to honor the module already constructed for the FFN branch.

Checklist

• I kept the change narrowly scoped.
• I did not commit secrets, API keys, large checkpoints, or generated cache files.
• I reused fixtures from data/test_cases/ or assets from data/benchmarks/ where practical.
• I updated docs when public behavior, install steps, or entry commands changed.
• I documented API/GPU/checkpoint/official-repo requirements or confirmed none are needed.
• I did not promote demo, contract-only, normalizer-only, API-blocked, or partial evidence to leaderboard validity.