extract_numbers.py: normalize_number matches units by substring, so 150bps becomes 150,000,000,000

Summary

normalize_number in skills/ib-check-deck/scripts/extract_numbers.py matches unit suffixes by substring containment rather than equality:

for unit_key in sorted(multipliers.keys(), key=len, reverse=True):
    if unit_key.lower() in unit.lower():        # <-- containment
        return base_value * multipliers[unit_key]

'B' is a key worth 1e9, and 'b' is a substring of 'bps'. So every basis-point figure is multiplied by a billion.

Reproduce

>>> from extract_numbers import normalize_number
>>> normalize_number('150', 'bps')
150000000000.0          # expected 150
>>> normalize_number('150', 'x')
150.0                   # correct
>>> normalize_number('150', '%')
150.0                   # correct

End to end:

>>> [(n.value, n.normalized) for n in extract_numbers("Spread widened 150bps.")]
[('150bps', 150000000000.0)]

bps is in the pattern's own unit alternation, so this is a unit the extractor is explicitly built to recognise. Every rate, spread and margin-delta figure in a deck is inflated by 1e9, and each one then enters find_inconsistencies as a revenue/margin-category value competing with real figures.

Cause and suggested fix

Containment was presumably chosen so USD_million matches million. But it makes every single-letter key a substring trap: 'B'→bps is the live one, and the same shape would bite any future unit containing t, b, m or k.

Match the unit exactly, after normalising case, and keep a separate explicit mapping for the composite forms the caller builds:

MULTIPLIERS = {
    'trillion': 1e12, 't': 1e12, 'tn': 1e12,
    'billion': 1e9,  'b': 1e9,  'bn': 1e9,
    'million': 1e6,  'm': 1e6,  'mm': 1e6, 'mn': 1e6,
    'thousand': 1e3, 'k': 1e3,
    '%': 1, 'bps': 1, 'x': 1,          # dimensionless — explicit, not by omission
}
key = unit.lower().removeprefix('usd_')
return base_value * MULTIPLIERS.get(key, 1)

Listing %, bps and x as explicit 1× entries is worth doing on its own: today they return the base value only because nothing matched, which is the same code path as an unrecognised unit.

Note

Found while verifying the fix for #338 — the negative half of that test ("genuine suffixes must still parse") is what surfaced it. Filing separately since it is a distinct defect, but I have a fix ready and can include it in that PR or a separate one, whichever you prefer.