[BUG] preliminary fix for incorrect setting of _pretrained_attrs in case __init__ already sets fitted attrs

[fkiraly]

Bug #10531 is triggered in case __init__ already sets fitted attrs, e.g., for initialization (the diagnosis in the issue is incorrect, see discussion below).

This PR adds a preliminary fix: only those attrs are added to self._pretrained_attrs that do not already exist once pretrain is entered.

A test for regression is also added.

For a final fix, an API design discussion needs to be reopened first.

[Copilot]

Pull request overview

This PR implements a preliminary fix to BaseForecaster.pretrain so that _pretrained_attrs only tracks attributes that first appear during pretrain, avoiding misclassification of fitted-style attributes that were already present when entering pretrain (e.g., attributes initialized in __init__). It also adds a regression test targeting the reported behavior in #10531.

Changes:

• Capture public trailing-underscore attributes present at pretrain entry and exclude them from _pretrained_attrs.
• Add a regression test ensuring NaiveForecaster.pretrain does not register pre-existing attrs as pretrained params.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
sktime/forecasting/base/_base.py	Excludes pre-existing public *_ attributes from being added to _pretrained_attrs during pretrain.
sktime/forecasting/base/tests/test_base.py	Adds regression coverage to ensure pretrain doesn’t misclassify pre-existing attrs as pretrained.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.