Carry metric() to the next SI prefix when rounding reaches 1000

[gaoflow]

Summary

metric() can emit a four-digit mantissa with a too-small SI prefix:

>>> humanize.metric(999.9, "V")
'1000 V'        # expected '1.00 kV'
>>> humanize.metric(999.99, "V")
'1000 V'        # expected '1.00 kV'
>>> humanize.metric(999999, "V")
'1000 kV'       # expected '1.00 MV'
>>> humanize.metric(0.0009999, "V")
'1000 μV'       # expected '1.00 mV'

This contradicts the function's own docstring, which states the prefix is

always chosen so that non-significant zero digits are required ... 1,230,000 will become 1.23M instead of 1230K

1000 V / 1000 kV are precisely the 1230K-style outputs it promises to avoid, so the documented contract is the oracle here.

Root cause

metric() selects the SI prefix bucket from exponent = floor(log10(abs(value))) on the raw value, then rounds the mantissa to the requested precision. When the value sits just below a power-of-1000 boundary, the rounded mantissa becomes 1000, but the prefix was already locked in from the pre-rounding exponent, and the code never re-checks whether rounding crossed into the next bucket.

Fix

After computing the mantissa and its display precision, if round(abs(mantissa), digits) >= 1000 and a higher prefix is available (exponent < 30), advance to the next prefix bucket and re-divide. The top of the SI range (quetta, Q, 10^30) has no higher prefix, so the exponent < 30 guard leaves it unchanged rather than indexing past the prefix table.

Tests

Added four cases to the test_metric parametrization (the carry cases above). They fail on main ('1000 V' != '1.00 kV', etc.) and pass with this change. The full tests/test_number.py suite passes (223 passed), so existing metric() output, including the high-range QF/qA cases and the exact power-of-1000 inputs, is unchanged.

Disclosure: I developed this fix with AI assistance, under my direction. I reviewed and verified the boundary behaviour (positive carry, the negative-exponent mV path, and the top-of-range quetta guard) and the regression run myself, and I am happy to adjust the approach.