fix: non-numeric age sentinels get their own categorical group

[propcgamer20-png]

_age_to_numeric() returned None for both a genuinely missing cell and free text with no embedded number at all (e.g. "unknown", "prefer not to say") - the age-banding loop then counted every None straight into null_count, indistinguishable from real missing data. Contradicted SPEC.md section 2's own documented "anything else: treat as categorical" rule for exactly this case.

$ python3 -c "
import pandas as pd
from faircode.profiler import profile
df = pd.DataFrame({'age': [25,30,45,'unknown','unknown','unknown','prefer not to say',22,33,41]})
d = profile(df)['dimensions'][0]
print(d['n_groups'], d['missing_pct'])
"
3 0.4 # (before) - 4 real responses vanish into missing_pct
5 0.0 # (after) - 'unknown' (3) and 'prefer not to say' (1) get their own groups

New _is_categorical_age_sentinel() (faircode/profiler.py) / isCategoricalAgeSentinel() (assets/profiler-engine.js) distinguish three cases per unparseable value: None/NaN (missing), a numeric value out of the valid age range like a -1/999 sentinel (also missing - matches this profiler's existing, tested behavior for range-invalid numeric sentinels), and free text with no embedded number at all (categorical). Only the third case gets its own group.

First attempt at this used pd.isna()/typeof-only-numeric as the distinguishing check and broke
test_negative_age_sentinels_are_missing_instead_of_an_elderly_group - a -1 sentinel isn't NaN, so it was wrongly routed to a new '-1' categorical group instead of staying missing. Caught by running the full profiler suite before committing; fixed by checking for an embedded number via the same regex _age_to_numeric()/ageToNumeric() already use, not just NaN-ness.

Verified Python/JS parity directly (not just via the test suite): built a CSV from the issue's exact repro and ran both
faircode.profiler.profile() and 'node scripts/engine-js.js profile' against it - identical n_groups, missing_pct, and group labels/counts on both engines.

Added two regression tests to tests/test_profiler.py: the issue's exact repro, and a mixed case (real None/NaN alongside a text sentinel) confirming missing_pct only counts the genuine blanks. Full suite: tests/test_profiler.py 38/38, tests/test_js_parity.py 21 passed + 1 skipped, full suite 401 passed/1 skipped/4 failed (the 4 are tests/test_xlsx_edge_cases.py, confirmed pre-existing and unrelated - SheetJS CDN blocked in this sandbox).

Known follow-up, out of scope here: _intersections()'s labelize() helper (used by compare(), not profile()) still routes these same sentinel values through ageBand()/None the old way for cross-tabulation purposes. The issue's own repro and acceptance criteria are profile()- only; left as-is rather than silently expanding scope.

Fixes #487

Summary

Type

• Audit
• Explainer
• Bug fix
• Other

---

Audit checklist

• I opened or linked a corresponding issue first
• The folder is named after the domain, not the dataset
• unfair.py includes protected attributes and prints the required output format
• fair.py removes protected attributes and identified proxy variables
• Both scripts use random_state=42 and an 80/20 train/test split
• Proxy variables were actually tested, not just guessed
• unfair.png and fair.png are included as PNG screenshots
• The dataset is public and accessible without login or payment
• The dataset file is included, or DATA.md is included if the file is too large
• README.md includes the new results row and audit section
• A notebook was added if the audit benefits from one

Before fairness gap:

After fairness gap:

Reduction:

Protected attribute(s):

Proxy variables dropped:

---

Explainer checklist

• The file is in explainers/ and uses lowercase hyphenated naming
• It includes a plain-language definition
• It uses a real example from this repo or a documented real-world case
• It includes runnable Python detection or measurement code
• It acknowledges limitations or trade-offs
• It links to related explainers or repo projects
• It includes 2-3 primary sources
• The Explainers table in README.md was updated

---

Linked issue

Closes #

[vercel]

@propcgamer20-png is attempting to deploy a commit to the yashkewlani2020-gmailcom's projects Team on Vercel.

A member of the Team first needs to authorize it.

[github-actions]

@yakew7 @ahmdkaml - new PR touching a path you own, please review.