fix(exif): FNumber prints through PrintFNumber, not the raw quotient

[swackhamer]

ExifIFD:FNumber disagreed with ExifTool in 1,035 of the 4,238 sample-corpus files: 4 where ExifTool prints 4.0, and 0.640234375 where it prints 0.64.

Root cause

Exif.pm 0x829d (line 1853) is PrintConv => 'Image::ExifTool::Exif::PrintFNumber($val)', and format_tag_value had no arm for it. The rational therefore fell through to the PrintConv-less-rational rule, which prints the quotient the rational reader already rounded -- RoundFloat($val, 10), i.e. %.10g. That is the right rule for a tag with no PrintConv and the wrong one here, in both directions: it drops the decimal place %.1f keeps (964 files), and prints the full stored expansion where ExifTool rounds (another 71).

PrintFNumber, Exif.pm:5715, reproduced exactly:

my $val = shift;
if (Image::ExifTool::IsFloat($val) and $val > 0) {
    # round to 1 decimal place, or 2 for values < 1.0
    $val = sprintf(($val<1 ? "%.2f" : "%.1f"), $val);
}
return $val;

Both halves of that condition are load-bearing:

• Below 1.0 it is two decimal places, not one. GPS.jpg stores 0.640234375 and ExifTool prints 0.64.
• $val > 0 guards the sprintf, so a zero is returned unformatted. CanonEOS20Da.jpg, NikonSUPER_COOLSCAN9000ED.jpg and SonyNEX-VG900.jpg store a zero FNumber and ExifTool prints 0, never 0.0.

Composites are untouched, by construction and by measurement

This is a display conversion and 0x829d has no ValueConv, so the one Composite that consumes FNumber -- Aperture (Exif.pm:4782, Desire => { 0 => 'FNumber', 1 => 'ApertureValue' }, ValueConv => '$val[0] || $val[1]', which applies PrintFNumber itself) -- must keep receiving the raw quotient.

It does. Composites are derived from the stored TagValue before format_tag_value runs at all (composite::lookup_key), and FNumber is deliberately not in apex_value_conv, the list of tags whose stored value is not what a reader wants. A test pins both halves of that.

Measurement

Per file, keyed Group1:Name, against ExifTool 13.59 over the 4,238-file corpus (oxidex reads 4,104 of them). JSON parsed with parse_float=str; files keyed on full relative path.

matched set	+1,030 / -0
files	1,030 improved, 0 regressed
keys changed	ExifIFD:FNumber, and nothing else -- no maker-note *:FNumber moved, because the arm is gated on the rational form the EXIF reader produces
Composite:*	0 values changed, across all 35 Composite keys and the 3,719 files carrying Composite:Aperture

The five files still short of ExifTool are pre-existing gaps this does not touch: CanonRaw.cr3 does not emit the tag, and four Samsung Anycall files store FNumber as a two-element rational array whose second element has a zero denominator (3 0 vs ExifTool's 3 undef) -- the array path, not this one.

Mutation test

Removing the arm fails fnumber_reaches_print_fnumber_from_its_rational with 4 for 4/1. Replacing the two-branch sprintf with an unconditional %.1f fails quotients_round_to_one_place_or_two_below_one (0.6 for 0.64) and non_positive_values_are_returned_unformatted (0.0 for 0).

🤖 Generated with Claude Code