fix(raw): decode Samsung linear DNGs correctly

[VailElla]

Note

These decoder changes have been resubmitted upstream as dnglab/dnglab#810. This downstream PR remains open temporarily because CyberTimon/RapidRAW#1380 currently pins commit ee86c32e3d85ff6057fdc8b3c3eed38923712d91. Once the upstream PR is merged and this fork is synchronized, RapidRAW's dependency will be refreshed and this PR will be closed as superseded.

Summary

• honor BlackLevelRepeatDim for interleaved linear DNG data instead of assuming the black- and white-level arrays have the same length
• parse JPEG DRI segments, validate RST0-RST7 ordering, resume entropy decoding, and reset the lossless JPEG predictor at restart boundaries
• reject truncated entropy segments instead of consuming synthetic zero-fill bits, and validate restart padding and trailing entropy bytes
• reject unsupported restart sampling in every component, truncated DRI segments, undeclared or zero-interval restart markers, invalid BlackLevel counts, and malformed normalization metadata instead of panicking or silently producing corrupted pixels
• warn and skip black-level correction when generic black/white-level array lengths still do not match, rather than aborting the process

Root cause and impact

This enables the decoder-side fixes needed by:

• CyberTimon/RapidRAW#777: the Samsung S23 Ultra sample has a valid 2x2x3 black-level repeat pattern plus three white levels. It also uses lossless JPEG restart markers; the previous bit pump stopped at the first marker, corrupting rows after the first restart interval.
• CyberTimon/RapidRAW#884: the Samsung Galaxy Z Fold6 JPEG XL sample uses the same valid repeated black-level representation, which previously reached a black/white-level length mismatch.

Both public issue samples now complete full RAW decoding. This PR intentionally does not implement default-rendering metadata such as BaselineExposure or OpcodeList2 gain maps; those belong to later rendering stages. RapidRAW#1380 applies baseline exposure, while GainMap rendering remains a planned follow-up and pixel parity is not claimed here.

Validation

• cargo +1.96.1 test -p rawler --test linear_blacklevel --test jpeg_restart (18 passed)
  • covers predictor reset, RST ordering, stuffed bytes, fill bytes, all-component sampling validation, undeclared and zero-interval restart markers, missing entropy, invalid padding, extra entropy, truncated final intervals, and malformed DRI metadata
• direct regression probe confirmed a generic black/white-level length mismatch returns without panic and leaves pixels unchanged
• cargo +1.96.1 check -p rawler
• cargo +1.96.1 build -p dnglab
• decoded the public Nikon ZR support dnglab/dnglab#777 and #884 attachments through dnglab analyze --raw-checksum
  • Nikon ZR support dnglab/dnglab#777: fedb82718ee20b5a309678872527cb75
  • #884: 51abe49e5950c6987ac347548c2ac487

The issue fixtures are used only for local validation and are not included in this branch.

[SimonIT]

Wouldn't it be better to submit the pull request to dnglab itself? This way, when Timon updates this fork, the changes will be pulled in automatically

[VailElla]

Thanks — agreed. I moved the decoder changes to a clean branch based on dnglab/dnglab:main and opened the upstream PR here: dnglab#810

I’ll keep this PR open temporarily because CyberTimon/RapidRAW#1380 currently pins this fork revision. Once the upstream PR is merged and this fork is synced, I’ll update the RapidRAW dependency and close this PR as superseded.