Never drop or shadow feature_dict groups with auto-filled columns (#81)

[upadhyan]

Fixes #81.

Root cause

_feature_dict_to_features stored user groups and auto-filled singleton columns in one dict keyed by logical name. An int key equal to the index of a column no group lists was overwritten by that column's auto-filled singleton, so the group was silently dropped from training:

configure_feature_dict(2, {0: [1]}).features  # [[0]] -- column 1 gone

A str key like "0" didn't lose data but produced duplicate logical names (("0", "0")), making export_text / plot_tree labels ambiguous.

Fix

Auto-filled column i is named str(i) unless a user key already has that name; then it becomes "<i>_1" (_2, … until free). User keys always keep their names.

configure_feature_dict(6, {0: [2, 3]}).logical_names
# ('0', '0_1', '1', '4', '5')   -- group "0" = [2, 3], column 0 = "0_1"
configure_feature_dict(3, {"0": [1, 2]}).logical_names
# ('0_1', '0')

Keys whose string forms coincide ({0: …, "0": …}) now raise ValueError, since their names would be ambiguous. Ordering is unchanged (int keys and auto-filled columns first, then str keys, each by name). Docstrings for configure_feature_dict and ProcessedFeatures.logical_names updated.

(The first commit raised ValueError on collision; the follow-up commit replaces that with renaming.)

Why no test caught it

• Every existing int-key test happened to list the key's own column ({0: [0, 1, 2]}, {i: [i]}, …), so the collision never occurred.
• Several regression tests compare against configure_feature_dict itself, so they share the bug.
• No test asserted "every column appears in exactly one logical feature".

Tests

• tests/test_features.py: rename cases (int key, str key, suffix already taken), same-string-form keys raise, fit keeps the group, and every-column-exactly-once + unique-names invariant across int/str/mixed keys.
• tests/regression/test_sgtlearn_features_py__feature_dict_to_features_1.py: 23 blind tests from test-checker run r-20261002-215910-e7c6 (0 misguided). Three that encoded the interim raise / duplicate-name behaviour were updated to the renaming rule.
• Full suite: 1276 passed, 2 skipped.

Not in this PR

• int keys still sort as strings (10 before 9), as documented; changing it would reorder feature_importances_.
• Tutorial categorical-features.ipynb cell 3 and the ProcessedFeatures docstring are also edited on the Update Stale Docs and Tutorials #77 docs branch; expect a small docstring conflict there.

🤖 Generated with Claude Code