Least-squares chapter: blocks that break when run top to bottom, and two claims that do not hold

[kgdunn]

Running least-squares-modelling/ as one script (the chapter's own contract) against process-improve 1.78.0:

Blocks that fail

• investigating-an-existing-linear-model.rst:185 sets pd.options.plotting.backend = "plotly". Two later blocks were written for the matplotlib backend and break once that option is set: summary-of-steps-to-build-and-investigate-a-linear-model.rst:153-158 (subset.plot.scatter(...).set_title(...): a plotly Figure has no set_title) and multiple-linear-regression.rst:230 (.plot(..., grid=True): plotly's line() rejects grid). Either use plotly idioms in those blocks (the book's stated style) or stop switching the backend mid-chapter.
• least-squares-exercises.rst:975-981: digester.hist(figsize=..., bins=30, color="lightblue") fails under the plotly backend (color must be a column), and digester.corr()["Y-Kappa"] raises ValueError: could not convert string to float: '31-00:00' because the Kamyr file's Observation column is text. The question at line 955 says to drop non-numeric identifier columns; the starter code never does.
• covariance-and-correlation.rst:158-180 is a fill-in-the-blank exercise (___) and cannot run; it needs a .. code-check: skip marker with the reason.

Claims that do not hold

• least-squares-models-with-a-single-x-variable.rst:269-271: "everything shown here carries over unchanged to scikit-learn's own LinearRegression". The block below reads mymodel.coefficients_ (sklearn has coef_, not coefficients_) and relies on print(mymodel) rendering an R-style summary; neither carries over. Only intercept_, fit, predict and score do.
• multiple-linear-regression.rst:286-288: "a single column of random noise moves R2 only in the fifth decimal (0.93831 to 0.93832)". Over 1000 seeds the median new R2 is 0.93849 (fourth decimal; expected increase (1-R2)/(n-k) = 0.00042); only 6% of seeds round to 0.93832. No seed is given, so the sentence is not reproducible as written.

Smaller items

• enrichment-topics.rst:130 and least-squares-models-with-a-single-x-variable.rst:268 link to github.com/kgdunn/process_improve (underscore, HTTP 403).
• least-squares-exercises.rst:894 calls a process_improve block "A Pandas / scikit-learn version".
• outliers-discrepancy-leverage-and-influence-of-the-observations.rst:108-110 says unconditionally that leverage_ and influence_ are provided; the library computes them only for a single predictor with an intercept and returns NaN otherwise (OLS.leverage_ and OLS.influence_ are silently NaN for multiple regression and for fit_intercept=False process-improve#545).
• least-squares-models-with-a-single-x-variable.rst:208-211: the comment shows const / x1 labels, but mod_ls.params on an ndarray prints an unlabelled array. :296: "the shape these estimators expect"; OLS.fit reshapes 1-D input itself.
• least-squares-exercises.rst:865: the upper bound 9.49 is written as 9.4.
• enrichment-topics.rst:148-152: comment says "slope ... and intercept", the print shows intercept first.

All 54 quoted numbers in these files reproduced except the fifth-decimal claim. Found by executing every code block of the book (#274).