fix(feature_utils): bind X and ndf_ on the routes that read them - #2097

Fifth of the pyright-ratchet cleanup series. Ratchet 255 → 245.

Two functions read a local that an earlier branch may never have assigned, and both are reachable.

transform()

y and T are pre-initialized at the top; X is bound only inside the nodes/edges branches. kind is a plain str that FastEncoder stores without validation, so any other value falls through to if not tX.empty and not X.empty:

transform(df, DataFrame([]), res, 'bogus', ...)
master: UnboundLocalError: cannot access local variable 'X'
here:   X.empty=True y.empty=True

get_numeric_transformers()

Returns ndf_ unconditionally but assigns it only under if ndf is not None — while the sibling y_ = y is assigned before its guard, which is the shape the author clearly intended. get_numeric_transformers(None, None) raised UnboundLocalError on master; it now returns ndf_=None.

Two sites deliberately left alone

Both looked like the same defect and neither is a bug:

• The skrub names are function-locals from a guarded import, but every use sits under the same has_skrub flag — including the isinstance at line 1025, which I checked individually.
• scale()'s scaler/scaler_target look unbound for an unrecognized kind, but kind is validated upstream (kind must be one of 'nodes' or 'edges') and a missing encoder raises AttributeError before the branch chain. I tried to construct a reaching call and could not.

Hardening those would have meant three # pragma: no cover lines in a file no gate-feeding lane can execute — a poor trade for changes that fix nothing. They stay in the ratchet baseline, visible.

A gap worth naming

feature_utils.py collects zero coverage in the lanes the changed-line gate combines (core, gfql, polars), because none of them installs scikit-learn — the umap/ai lanes that do exercise it produce no coverage artifact at all. Any change to this file is therefore invisible to that gate.

The one sklearn-only line here carries a pragma, and I checked its claim rather than asserting it: line 873 is covered when the umap tests run with sklearn present.

Wiring an umap/ai coverage artifact into changed-line-coverage would fix this properly and unblock future work in this file. It is a real CI change (new artifact + combine step + a needs edge, and per #2018 a skipped prerequisite skips the gate), so I have not folded it in here.

Verification

bin/lint.sh (ruff, type-hygiene 4448, comment-density 1151) · pyright guard 245 · mypy 351 files · changed-line coverage 1/1 = 100% · tests pass with and without sklearn (2 passed/1 skipped, then 3 passed) · 9 passed in the umap env across feature/umap suites · meta-guards run · the recognized-kind test confirmed discriminating by disabling the nodes branch.

🤖 Generated with Claude Code

https://claude.ai/code/session_017ropeBMLJUuy6ViYwy15ud