[initial-d]

Summary

The training dataset previously coupled today's forward-label validity to tomorrow's factor availability. A limit-move exclusion can invalidate tomorrow's factors while today's observed return label remains valid. FactorDataset accepts a separate already-aligned label_mask, and the CLI supplies target_01's mask. Calls without the new argument retain the legacy intersection. Nonfinite training inputs are excluded and mask/label shapes are checked.

Forward labels now require finite positive closes at both endpoints and a finite computed return. Zero/negative prices no longer become clamped extreme returns; masked NaN/Inf produce zero values with false masks. Genuine zero returns stay valid. Positive integer horizons at or beyond panel length have no valid labels; multi-row horizons check endpoints only.

The mask contract distinguishes observations, factor exclusions, historical labels and order feasibility. This can change training sample counts; it is not new market-performance or alpha evidence.

Type of change

• Bug fix
• Documentation
• Test or benchmark

Reproducibility

• Deterministic mask/label/alignment/nonfinite/endpoint regression tests.
• Updated API and research-policy documentation.
• No external market data, private data or credentials.

Collaboration

• Solo contribution.

Validation

Clean Python 3.12.14 / PyTorch 2.14.1, macOS 26.6.2, Apple M5 CPU:

python -m pytest: 173 passed
python -m ruff check src tests: All checks passed

docs/audits/training-masks-20261023/ preserves the original checks and older-environment failures. docs/audits/forward-labels-20261024/ records the endpoint follow-up and raw test output. All evidence is synthetic/unit-test engineering evidence; archived public-data returns are not revalidated.