faircode/proxy.py's _labelize() silently drops age-sentinel rows from the chi-squared proxy test

[yakew7]

Where: faircode/proxy.py's _labelize() (used by the proxy-hints chi-squared test) vs. faircode/profiler.py's _intersections(), which already special-cases this exact scenario.

The gap: _labelize() claims in its own docstring to be "the same value normalization the intersection crosstab uses (age → bands)" - but it isn't. For an age column containing free-text sentinels with no embedded digit (e.g. "unknown", "prefer not to say"), _age_to_numeric() returns None, then _age_band(None) also returns None, and pd.crosstab() silently drops every row carrying that value from the chi-squared test - exactly the bug profiler.py's _intersections() already had fixed for issue #524 (that fix's own commit message: "otherwise labelize() maps them to None and pd.crosstab silently drops those rows"). _labelize() in proxy.py never received the equivalent fix.

Repro:

import pandas as pd
from faircode.proxy import _labelize

df = pd.DataFrame({
    'age': ['25','30','unknown','40','unknown','35','28','unknown','50','unknown',
            '22','60','unknown','33','unknown','45','29','unknown','38','unknown'],
    'gender': ['M','F']*10,
})
lab = _labelize(df, 'age', 'age')
ct = pd.crosstab(lab, df['gender'])
print(ct.to_numpy().sum(), 'of', len(df), 'rows survived')
# 14 of 20 - all 6 "unknown" rows silently vanished

Impact: the proxy-hints chi-squared test (faircode profile --proxy-hints) silently understates its sample size and can bias the reported p-value/Cramér's V for any dataset with digit-free age sentinels - with no warning that rows were dropped. Not currently triggered by any dataset shipped in this repo, but reachable by any real-world one.

Fix direction: either import/reuse profiler.py's categorical-age-sentinel handling directly, or port the same fix to _labelize().