[ChrisW09]

Fixes #541

Problem

build_model() replaces the task model with freshly initialized weights. On an already fitted estimator it kept is_fitted_ = True, so predict(), evaluate(), save() and InferenceModel.from_estimator() (and a classifier's score()) silently treated the untrained network as the fitted model, while a regressor's score() loaded the previous fit's best checkpoint into it. This happened for classifiers and regressors (_FitMixin._build_model) and for LSS estimators (SklearnBaseLSS.build_model). fit() already resets the flag before rebuilding.

Fix

• A new _FitMixin._mark_unfitted() resets is_fitted_ to False when it was set, and clears _best_model_path, the previous fit's best checkpoint. A never-fitted estimator gains no is_fitted_ attribute, so profile() still leaves it unchanged.
• _FitMixin._build_model and SklearnBaseLSS.build_model call it. The classifier's build_model calls it before replacing classes_, so a build that fails on its labels (e.g. a y_val label unseen in y) leaves an unfitted estimator rather than one with the new classes and the old network.
• fit() sets the flag back to True when training completes, including fit(rebuild=False) after build_model().
• The Returns sections of the classifier, regressor and LSS build_model docstrings state that the built estimator counts as unfitted (predict() raises NotFittedError) until fit() completes.

Tests

tests/integration/test_build_model_state.py (new file):

• test_regressor_is_unfitted_after_rebuilding: predict() raises NotFittedError and _best_model_path is cleared after build_model(), and fit(rebuild=False) makes the estimator fitted again.
• test_lss_is_unfitted_after_rebuilding
• test_classifier_is_unfitted_when_rebuilding_fails: a build_model() that raises on an unseen y_val label leaves the classifier unfitted.

All three fail on main and pass with this change. The fix for #452 cleared the same stale flag for a failing fit().

Docs

Docstrings as above. docs/core_concepts/config_system.md already describes build_model() as building the network "without training its weights".

Checks

• just test: 3239 passed, 46 skipped, 28 xfailed; coverage 91.39%
• just docs: build succeeded (warnings treated as errors)
• just check: all hooks passed (ruff, prettier, pyright)
• The tests live in their own module so this branch merges cleanly with the other open PRs that touch fit.py, lss_base.py or test_fit_lifecycle.py (fix(models): respect enable_model_summary in fit #486, fix(models): honour ObservabilityConfig in LSS fit #500, fix(models): score the in-memory model without reloading the checkpoint #504). I checked this against all open PR branches.