[Yushu-Li]

I fixed patience-based early stopping when convergence_threshold is zero.

With early_stopping_patience = 3, convergence_threshold = 0.0, and a constant score of 0.0, the controller treats every tie as an improvement because 0.0 >= 0.0. It resets patience on every iteration and runs through the full iteration budget instead of stopping on the plateau.

Require the score to increase before applying the existing inclusive minimum-delta check. Equal scores then count toward patience at a zero threshold, while an improvement exactly equal to a positive threshold still resets it.

Four new controller regressions cover:

• A constant zero score stops after iteration 4 rather than running all 8 iterations.
• Tiny improvements to negative scores still reset patience with a zero threshold.
• An improvement exactly equal to a positive threshold still resets patience.
• Event-based stopping at a zero target still works.

The zero-threshold plateau regression fails on the original controller; the three compatibility controls pass.

Validation on macOS / Python 3.12:

• All 18 focused early-stopping tests passed.
• OPENAI_API_KEY=test-for-offline-tests PYTHONPATH="$PWD" python -m unittest discover -s tests -p 'test_*.py' -v: 958 tests passed.
• Black checks on both changed files, isort on the edited test module, and git diff --check passed.

I used AI assistance for implementation, regression tests, and local validation. No paid model calls or model-backed integration runs were performed.