[initial-d]

IC/RankIC training previously shuffled stock-days from different dates into the same correlation batch. Between-date level shifts can make the pooled correlation positive even when every daily cross-section has reversed stock ordering. The time split could also put stocks from the same date on both sides, and a training set smaller than batch_size silently performed no optimizer steps.

The trainer now defaults to one complete observed date per batch for ICLoss/RankICLoss, shuffles dates, and averages daily losses equally. Time validation holds out complete latest dates; val_split is a date fraction in this mode. Random daily splits also keep dates intact. Ordinary regression retains sample batching and includes the final partial batch. batch_mode="date" supports custom daily objectives; explicit sample batching is rejected for the built-in IC losses. Validation runs without autograd.

The PR also guards zero-variance operands before Pearson square roots: constant/singleton batches contribute zero loss and zero gradients instead of poisoning optimizer state with NaNs. Ordinary finite nonconstant batches retain the previous correlation formula/gradients. Correlation inputs must be aligned, nonempty and finite; gamma/temperature must be finite positive values.

Behavior and limits:

• Date mode ignores batch_size; memory depends on the largest cross-section. Soft RankIC still uses quadratic pairwise rank memory.
• Constant/singleton dates are retained with zero loss/gradients, so daily loss is an optimization diagnostic, not a mean of only defined IC values.
• Random splits remain reproducibility options, not chronological validation evidence. Longer label horizons still require a suitable purge gap.
• Validation boundaries and retained partial batches can change historical model diagnostics. No archived market result is revalidated and no alpha/investment claim is made.
• Based on main plus the prior gradient fix in this PR; excludes pending Separate label validity and reject unusable return endpoints #80/Add streamed daily factor screening and stable rolling moments #81.

Local verification for head c955c1a8b937cdf78a33b1ed7cedb3ddf16e0a33, Python 3.12.14 / PyTorch 2.14.1, macOS 26.6.2, Apple M5 CPU, deterministic unit/synthetic fixtures only (no external market data, no trading-cost assumption applicable):

• /tmp/projects-push-20261028-venv/bin/python -m pytest tests/test_cross_section_training.py tests/test_trainer_split.py tests/test_losses.py — 69 passed.
• /tmp/projects-push-20261028-venv/bin/python -m pytest — 209 passed.
• /tmp/projects-push-20261028-venv/bin/python -m ruff check src tests — passed.
• git diff --check — passed before commit.

Regression tests cover the pooled-correlation counterexample, ragged and singleton dates, date-disjoint time/random splits, equal-date validation weighting, actual updates on undersized regression datasets, four-dtype degenerate gradients, normal gradient parity and Adam recovery.

GitHub CI passed on the exact head above: Python 3.9/3.10/3.11 lint, full tests and CLI smoke checks. The maintainer explicitly requested direct merge on 2026-10-28 after these checks passed. Merged using the repository's permitted administrator override; no independent review is claimed and branch protection settings were not changed.