Fix reject-inference.md Script Execution Output to match its own code - #567

Closes #522.

The Script Execution Output block and the paragraph after it were pasted from an earlier version of the detection code and never regenerated. Running the file's own python block directly (seeded, byte-identical across repeated runs) gives:

                       Population AUC  Population Accuracy  Older Approval Rate  Younger Approval Rate  Age Fairness Gap
Naive (Approved Only)          0.9460               0.8851               0.5413                 0.5366            0.0047
IPW Reweighted                 0.9456               0.8813               0.5514                 0.5453            0.0062
Soft Parceled                  0.9238               0.8852               0.5413                 0.5369            0.0045

not the claimed 0.7812 -> 0.8345 AUC and 15.80 -> 2.30 point gap.

Why regenerating the prose (not just the table) was necessary

The old narrative ("15.80-point gap corrected down to 2.30") is not reachable with this simulation, not merely stale:

• y_true is generated with no age term - older and younger applicants repay at the same ~55% rate.
• Every model trains on credit_score and income_k only, both drawn independently of age_young.

So selection acting on age alone is ignorable for estimating P(Y | X); a correctly specified RandomForest recovers a near-parity score distribution with or without IPW, and there is no naive-model gap for the correction to close. IPW/parceling only bite under MNAR selection or when the outcome model omits a feature that drives selection - neither of which this minimal simulation creates.

Change

• Output block replaced with the real deterministic figures.
• Analysis rewritten to describe what the code actually shows: the disparity is entirely in the selection gate (older approval ~50%, younger ~35%; young applicants are 34.6% of the population but 27.1% of the approved pool), which no audit run on model scores or approved-only rows can see - i.e. exactly Why It Matters item 2.
• Added a one-line caveat that a seeded RandomForestClassifier's low-order digits can shift across CPU/BLAS backends (same cross-platform effect discussed in model-drift.md cites a fabricated single-run gap/p-value for unfair.py (6.39%/p=0.348 instead of the real 7.16%/p=0.2564) #521).
• Regenerated reject-inference.html, faircode/_explainers/reject-inference.md, and the llms-full.txt section.

No code in the fenced block was changed. check_em_dash.py, check_broken_links.py, check_generated_files_current.py all pass.