[PatrickHenkel1]

Fixes from the review of the release candidate (after #90), plus a one-time formatting of the whole code.

Fixes (466eee7)

1. AgentLib-MPC export of isolated lags. check_agentlib_compatibility now rejects lag groups that do not start at lag 0 (e.g. ['u', 'T_lag1']). AgentLib-MPC accepts such configs without an error, but only counts the inputs of a feature and passes its values from lag 0 on: for T_lag1 the MPC would feed T(k) instead of T(k-1), and for u_lag1 it would pass u, u_1 (shape mismatch at runtime). physXAI itself still trains with isolated lags (e.g. a dead time u_lag3); only the export needs the complete group. A contract test documents how AgentLib-MPC translates these configs.
2. ModularANN.online_pipeline with warm start. With a normalized px.Linear holding data-only terms (warm start on by default), the online pipeline crashed (AttributeError, no build context). If it had run, the warm start would have overwritten the loaded weights. A loaded model is now trained further from its own weights.
3. PCNN parameters after online training. The trained rates, consistency_margin and td.physics['initial_rates'] were computed with initial rates estimated again from the current data, not with the ones built into the loaded model. Reproduced: 3.72 K/h reported, 1.24 K/h in the model. They are now read from the model's constant_pcnn_scale_<j> layers.
4. Monotone networks need a hidden layer. With n_layers=0, CMNNModel, PINNModel and px.MonotoneMLP only had the non-negative weights of the output layer, so the prediction was increasing in every input, also with monotonicity -1. n_layers >= 1 is required now (also checked in the constructor, before the preprocessing).

CHANGELOG entries for 1 and 4; 2 and 3 concern features that are new in 1.0.

Formatting (dca1edb, 48249ef)

• ruff format (0.17.4) on the package, tests, examples and the code block in the README. Only formatting, ruff check is still clean.
• quality.yml also runs ruff format --check .. The ruff version is pinned there and in .pre-commit-config.yaml (v0.17.4), since the formatting can change between ruff versions. dev extra: ruff>=0.17.4.
• .git-blame-ignore-revs lists the formatting commit; GitHub skips it in blame automatically (locally: git config blame.ignoreRevsFile .git-blame-ignore-revs).

The formatting commit touches 50 files: open PRs based on refactoring_v1.0.0 will conflict and are easiest to reformat before merging.

Checks (local, Windows, Python 3.12)

• pytest: 272 passed, 1 skipped (RNN contract test, AgentLib-MPC main without RNN support)
• tests/smoke_examples.py: 13/13 examples OK
• ruff check physXAI tests examples, ruff format --check .: clean
• Sphinx (sphinx-apidoc + sphinx-build -W, Sphinx 6.2.1): no warnings

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.