[ozenyunus35]

Fixes #15574

With --to=si, --to=iec or --to=iec-i, a value printed without a suffix did not use --round:

• The value was printed with format_float(i2, 0), which rounds half to even. So --to=si 12.1 gave 12 instead of 13, the default from-zero result.
• consider_suffix checked the unrounded value against bases[1] - 1.0. So --round=down 999.9 got a suffix and printed 0.9k, where GNU prints 999.

consider_suffix now rounds the value with the method first and leaves it without a suffix if the rounded value still fits. It uses div_round, or round_with_precision(.., 0) when a --format precision is given, matching how each case is printed later, and returns that rounded value, so transform_to does not round it again. Values below 10 keep one decimal and are then printed with ties to even, as before. GNU does the same: with the default method, 1.45 gives 2 and 2.5 gives 2, and the results from #12032 (1.4 and 0.6 give 1) stay the same.

export LC_ALL=C
numfmt --to=si 12.1 99.1
# before: 12 / 99
# after:  13 / 100    (same as GNU 9.12)

numfmt --to=si --round=down 999.9
# before: 0.9k
# after:  999         (same as GNU 9.12)

numfmt --to=iec --round=down 1023.9
# before: 0.9K
# after:  1023        (same as GNU 9.12)

I added test_round_without_suffix, which checks all five methods with --to=si, --to=iec and --format=%.1f. It fails without the change, and all test_numfmt tests and the uu_numfmt unit tests pass.

I also compared 90 option combinations with GNU 9.12 in the C locale: --to=si/iec/iec-i, each --round method, and no format, --format=%f/%.0f/%.1f/%.2f and --padding=8. The input was 36 values between -999.9 and 999999, including values near 10, 999 and 1023. Before the change most of them differed. After it the output is the same, except for one input: -0.6 with --round=up or towards-zero and a --format precision prints -0 (or -0.0), where GNU prints 0. That is already the case on main and is not changed here.

cargo bench -p uu_numfmt shows no change against the merge base.

I used an AI assistant to help draft parts of the code and tests; I reviewed the changes and ran the tests myself.

[sylvestre]

why not return unscaled here? then transform_to wouldn't need to round it again

[ozenyunus35]

@sylvestre Thanks! consider_suffix now returns unscaled for the no-suffix case, so transform_to prints it without rounding again. Below 10 the two steps (one decimal, then ties to even) are what GNU 9.12 does too: with the default from-zero, 1.45 gives 2, 2.5 gives 2 and 0.5 gives 0, and uutils now gives the same. I also added a --format=%.1f case for all five methods to test_round_without_suffix.

[sylvestre]

below 10 this rounds twice now (one decimal, then ties-even), so 1.45 gives 2 but 2.5 gives 2, is that what GNU does?
could you please add 1.45 / 2.5 / 9.95 to the test?

[sylvestre]

please add a --format=%.1f case too, the precision branch isn't covered