A few small follow-ups from recent PRs that mostly fall in the "good first issue" bucket. I'm opening them here as a single umbrella so they can be claimed and worked in parallel instead of buried in the individual PR threads.

1. Stale docstring in TreePredictor.deserialize

src/gluonts/ext/rotbaum/_predictor.py around line 367:

def deserialize(cls, path: Path, **kwargs: Any) -> \"TreePredictor\":
    \"\"\"
    ...
    It loads the predictor class with the serialized arguments. It then
    loads the trained model list by reading the pickle file.
    \"\"\"

After #3176, the method reads model_list.json, not predictor.pkl. The docstring should match. (#3294 touches this line as part of its fallback change; if that PR lands first this item is resolved.)

2. Regression coverage the recent PRs didn't add

Each of these PRs changed an observable behaviour but didn't pin it in a test; the list is short enough that each one is a small, self-contained test file:

• Recover from a checkpoint if training failed with an exception #3265 — checkpoint recovery: no test that trainer.fit raising a ValueError on a tiny model leaves best_model_path populated and the returned predictor usable. Without this, the broad except Exception discussed in Torch estimator: broad except Exception in checkpoint recovery masks unrelated failures and loses optimizer/RNG state #3295 cannot regress visibly.
• Make observed_value_field optional in TFTInstanceSplitter #3259 — observed_value_field: Optional[str]: no test passing None through TFTInstanceSplitter on a toy dataset. The happy path with the default field name is covered; the new None branch isn't.
• Allow disabling scaler for DL models #3251 — scaling: Optional[str]: no test exercising scaling=None (and, after Torch: reject unknown scaling values instead of silent NOPScaler #3293, no test asserting that unknown scaling strings raise ValueError). A single parametrize across patch_tst, lag_tst, d_linear, tide would cover all four estimators.
• Fix freq string issues in datasets #3232 — freq alias migration: the parametrize in test/dataset/test_tsf_reader.py pins the mapping, but there is no test asserting that the emitted strings don't trigger FutureWarning on a supported pandas (i.e. no contract test against pandas 2.x). Datasets: finish pandas 2 freq alias migration in _tsf_reader #3291 fixes the values; a follow-up test that fails loudly if anyone re-adds \"M\"/\"Q\"/\"Y\"/\"S\" would be worth adding alongside.
• Enable QuantileOutput for TiDE model #3189 — TiDE QuantileOutput: no test instantiating TiDEEstimator(distr_output=QuantileOutput(...)) through create_lightning_module(). The existing tests exercise StudentTOutput() only.
• PatchTST: Add support for time features #3167 — PatchTST time features: no test asserting shape/ordering of past_time_feat/future_time_feat when num_feat_dynamic_real > 0. Current tests keep it at the default 0.

Not asking for anything cosmetic

Each bullet above was surfaced while validating a substantive behaviour change; none are style nits. Happy to send the docstring fix now and the test additions as separate small PRs if maintainers would like them split up by topic. Flagging good first issue since the scope per bullet is small and the relevant PR already explains the intended behaviour.