Add DNG, CR2, NEF, and ARW RAW format support

[tonimelisma]

Closes #59.

Summary

Adds DNG, CR2, NEF, and ARW RAW format support. All four are TIFF containers — imageDecoderRAW validates the TIFF header then hands off to the existing EXIF/IPTC/XMP decoders. All four pass golden comparison against exiftool.

Context-dependent tag naming

The RAW decoder itself is straightforward. The hard part was that exiftool gives the same tag ID different names depending on IFD, Compression, and SubfileType. Without replicating this, CR2 and ARW golden tests fail.

The switch statements in metadecoder_exif.go (lines 354–425) are derived from exiftool's lib/Image/ExifTool/Exif.pm. We implement the subset relevant to our formats:

Tags 0x0201/0x0202 (exiftool has 10 condition entries, we implement 4):

Our condition	Exiftool equivalent	Result
dirName == "IFD1"	Entry 1	ThumbnailOffset/Length
dirName == "SubIFD"	Entry 6	JpgFromRawStart/Length
HasPrefix(dirName, "SubIFD")	Entries 8–9	OtherImageStart/Length
default (IFD0, etc.)	Entry 10 (fallback)	PreviewImageStart/Length

Tags 0x0111/0x0117 (exiftool has 7 condition entries, we implement 5):

Our condition	Exiftool equivalent	Result
comp == 34892	Entry 2	OtherImageStart/Length
comp == 52546	Entry 3	PreviewJXLStart/Length
format == CR2 && dirName == "IFD0"	Entry 5	PreviewImageStart/Length
(DNG|TIFF) && comp == 7 && sft != 0	Entry 6	PreviewImageStart/Length
default	Entry 4	StripOffsets/ByteCounts

Skipped entries are for formats we don't support (MRW, NRW, SRW, RIFF/MOV). currentCompression and currentSubfileType are tracked per-IFD and reset at each decodeTags() call — safe because TIFF requires ascending tag order, and both precede the offset tags.

4 of 9 branches are exercised by golden tests (CR2, ARW, DNG, NEF). The other 5 lack test files — they require lossy DNG, JPEG XL DNG, or newer Nikon NEFs where the smallest CC0 samples are 22–94 MB. These branches are straightforward value comparisons directly translated from exiftool.

Other changes to shared code

• helpers.go — convertNumbersToSpaceLimited returns single values as-is instead of "". Fixes single-value LONG tags (e.g. BlackLevel).
• gen/main.go — identify failures now non-fatal (ImageMagick can't read RAW without libraw delegates). Golden infra already handles missing .config.json.

Test images

CC0 samples from raw.pixls.us, smallest available per manufacturer. Total ~32 MB:

File	Camera	Size
sample.dng	Canon PowerShot SD450	6.3 MB
sample.cr2	Canon PowerShot S90	13 MB
sample.nef	Nikon D1	3.9 MB
sample.arw	Sony DSLR-A330	9.2 MB

Is ~32 MB of test data acceptable, or would you prefer I find smaller samples? Cropping/downsampling risks altering IFD structure.

Tests

• TestDecodeRAW / TestDecodeRAWConfig — EXIF and CONFIG assertions for all four formats
• Golden comparison passes for all four (none in goldenSkip). "Binary data" tags skipped.
• Fuzz tests and benchmarks for all four formats

[tonimelisma]

To clarify: I don't claim to understand all the magic behind exiftool's parsing behavior but to your point, there must be some hard-earned wisdom there so I saw it best to replicate the logic.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 76.26263% with 47 lines in your changes missing coverage. Please review.
✅ Project coverage is 78.16%. Comparing base (69db29f) to head (ea79203).

Files with missing lines	Patch %	Lines
imagedecoder_raw.go	77.96%	35 Missing and 4 partials ⚠️
metadecoder_exif.go	53.84%	4 Missing and 2 partials ⚠️
gen/main.go	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #61      +/-   ##
==========================================
- Coverage   78.25%   78.16%   -0.10%     
==========================================
  Files          15       16       +1     
  Lines        1890     2079     +189     
==========================================
+ Hits         1479     1625     +146     
- Misses        279      317      +38     
- Partials      132      137       +5     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[bep]

To clarify: I don't claim to understand all the magic behind exiftool's parsing behavior but to your point, there must be some hard-earned wisdom there so I saw it best to replicate the logic.

I don't understand all of it either, and I certainly don't agree with all of his choices. But there's som much odd image metadata out in the wild, so without any reference output, I'd be lost.

30 MB test data is too much. Especially considering the relatively poor test coverage in this PR (for the exif changes, that is, see above), which means that we either have 1. Too few test files or 2. Too much code. I have not thought about this too hard, but my initial thinking is that I'm not convinced about how useful metadata information about preview images is for us.

[tonimelisma]

Pushed a follow-up commit:

Removed context-dependent tag naming (~100 lines deleted from metadecoder_exif.go). The switch blocks that renamed tags based on IFD context / compression / format are gone, along with the currentCompression/currentSubfileType state tracking. Tags 0x0201/0x0202 now always use their default TIFF names (ThumbnailOffset/ThumbnailLength) and 0x0111/0x0117 always use StripOffsets/StripByteCounts. Patch coverage on decodeTag goes from 55% to 91%.

Truncated test images to metadata-only (32 MB → 275 KB). The strip/tile image data is never read by the metadata decoder, so truncating past the IFD entries is safe. Golden files are unchanged.

SubIFD following (0x014a) is kept — without it, DNG/NEF/ARW report IFD0 thumbnail metadata (e.g. ImageWidth=128) instead of the main raw image.

Golden test results after this commit:

• DNG and NEF pass golden comparison (not in goldenSkip).
• CR2 is in goldenSkip — it chains 4 IFDs via next-IFD pointers (not SubIFDs), so we only reach IFD0/IFD1.
• ARW is in goldenSkip — exiftool renames IFD0's 0x0201/0x0202 to PreviewImageStart/Length; our default names collide with IFD1's ThumbnailLength.

Both are still exercised by TestDecodeRAW and TestDecodeRAWConfig.

[bep]

Thanks, this looks great. I will wrap this up in #62.

As to test data, I have thought about pulling in some extended data sets for CI builds (and development), which I guess would also make some of the licensing questions easier to reason about, but that needs to wait for another day.