fix(binning): learn bin edges at fit instead of per transform call

[ChrisW09]

Fixes #11.

Problem

fit recorded only n_features_in_ and total_output_dim_. transform called
pd.cut(X.squeeze(), bins=bins_spec, retbins=True) on whatever batch it was handed and used
those edges — so the encoding depended on the transform batch, not on the training data:

t = CustomBinTransformer(output_dim=4).fit(np.linspace(0, 10, 100).reshape(-1, 1))
t.transform([[0], [2.5], [5], [7.5], [10]])           # [0 0 1 2 3]
t.transform([[0], [2.5], [5], [7.5], [10], [100]])    # [0 0 0 0 0 3]

Same five rows, different codes — and neither result reflects the 100-row fit. This is silent
train/serve skew: no error, and the codes look entirely plausible.

transform also rejected fewer than three rows, so single-row inference was impossible.

Fix

• Compute edges in fit, store as bin_edges_ (one array per column), reuse verbatim.
• Clip into the fitted range at transform. pd.cut returns NaN outside its edges, which
  would put NaN into an otherwise integer column; clipping puts unseen extremes in the
  first/last bin instead. Flagged as an open question in the issue — this is the choice I
  made, easy to change to a reserved code if you'd rather.
• Move the numeric-dtype guard to fit, alongside the edge computation, so string input is
  rejected at fit rather than at first transform.
• Drop the X.shape[0] <= 2 guard from transform. It only ever mattered for deriving an
  equal-width range, which now happens at fit — where pandas handles small inputs by
  widening the range.

Result

transform(rows)              : [0 0 1 2 3]
transform(rows + outlier)    : [0 0 1 2 3 3]     <- first five unchanged
transform(single row [5.0])  : [1]
Preprocessor [0, 5, 10]      : [0 2 4]
   with an outlier appended  : [0 2 4 4]         <- was [0 0 0]

One existing test changed

test_custom_bin_transformer_raises_on_invalid_shape asserted that
transform(np.array([[0.1]])) raises "Input must have more than 2 observations." — i.e. it
encoded the behaviour this issue asks to remove. I repurposed it to cover the real shape
guard (a 2-column input raises PretabDataError) and added
test_custom_bin_transformer_raises_before_fit for the NotFittedError path that the
un-fitted call now takes.

No other existing test needed touching. Two tests that fit on a single row still pass,
since pandas widens the range for degenerate input.

Tests

Seven added: edges learned at fit, codes independent of batch composition, codes reflecting
the fitted range, single-row transform, out-of-range clipping, explicit-edge round-trip, and
string input rejected at fit.

Full suite: 488 passed, 9 xfailed. pyright unchanged at 71. Ruff on the touched test file
goes from 4 pre-existing errors to 3 — the RUF043 disappears because the test I rewrote now
uses a raw-string match=.

🤖 Generated with Claude Code