[Bug]: Zero turnover limit liquidates, CAE reports convergence without training, SAE BatchNorm crash on one-row batch, attention NaN/wrap edge cases #55

Affected library

models

Affected package version

0.1.4 (main at 786ca1b6)

Environment

Linux, source checkout of main at 786ca1b6 with the repo's dev environment (uv sync); offline test suite.

Observed behavior

1. turnover_limit=0 liquidates the previous book.
2. CAE with hidden units returns converged=True when every minibatch has one row and no step runs.
3. SAE crashes in BatchNorm when the last minibatch has one row.
4. Graph attention returns NaN outputs/gradients for a node with no neighbors.
5. Cross-sectional attention wraps the slice when lag > window.

Expected behavior

1. Zero holds the previous weights; negative/non-finite limits raise. 2. The fit raises. 3. Training completes. 4. Outputs stay finite. 5. Keys are all-pad.

Minimal reproduction

Each case has a focused unit test on quantum-like/models branch review/fixes (one commit per fix, on top of 786ca1b6). Every test fails on main and passes with its fix:

• tests/test_financial_correctness.py::test_zero_turnover_limit_holds_the_previous_book, ::test_negative_turnover_limit_is_rejected
• tests/test_cae.py::test_cae_rejects_fits_that_never_update_parameters
• tests/test_sae.py::test_sae_trains_when_the_last_minibatch_would_be_a_single_row, ::test_sae_rejects_a_single_training_row
• tests/test_financial_correctness.py::test_graph_attention_stays_finite_when_a_node_has_no_neighbors, ::test_graph_attention_ignores_a_masked_only_neighbor
• tests/test_financial_correctness.py::test_cross_attention_lag_longer_than_the_window_does_not_wrap

Compatibility impact

Public behavior or results

Acceptance criteria

Each case has a unit test that fails on main and passes with the fix.

I have a PR ready with these fixes and will link it here.

Amir