[gmipf]

Fixes #1.

Two problems in the inflater made MS-ZIP blocks decompress wrong. Each commit fixes one and adds a test that fails without it.

Short codes at the end of the input. InflateCodes waited for as many bits as its lookup table uses before it decoded a code. An MS-ZIP block is a deflate stream on its own, so when a block ended with fewer bits than that, its last match was not decoded and the block came out short. This is what #1 reports: block 43495 of AgainstRome1.cab decodes to 32638 bytes and block 7 of folder 177 of Data2.cab to 32743 bytes, the same numbers as in the issue. Since 4dc1378 the missing end is filled with zeros, which only gives the right data when the block ends in zeros. The short blocks of AgainstRome1.cab and Group2.cab do, so main extracts those two correctly. The one in Data2.cab does not, so _86261BAE7A6B4C45A5D691D3320D5993 from the issue still comes out wrong on main. When no more input is there, the code is now looked up with the bits that are left and used if it is not longer than those.

A full 32 KiB history. SetDictionary kept only the last 32767 bytes of the previous block, so a match at the start of a block that copies from 32768 bytes back read a byte that was not in the window. None of the three disc cabinets below has such a match. Generated blocks with them decode the same in zlib, cabextract and 7-Zip, and wrong on main.

Measured on Linux with the cabinets from the issue, taken from the dumps of GameStar 07-2008 (Germany) (DVD-XL), Test Drive Unlimited (Germany) (En,Fr,De,Es,It) (Rerelease) and GameStar 09-2008 (Germany) (DVD-XL):

• Every MS-ZIP block of AgainstRome1.cab, the Data.cab set and Group2.cab (45124, 230187 and 14497 blocks) now decodes to the same bytes as zlib. On main, 6 of them come out short.
• ExtractionTool from SabreTools.Serialization main extracts the same files as cabextract with this branch. On main, only _86261BAE7A6B4C45A5D691D3320D5993 differs.

The split block in the spanned Data.cab set, which threw the exception in the issue, is already joined correctly on SabreTools.Serialization main.

Prepared with AI assistance (Claude Opus 5.5) and reviewed before submission.

[mnadareski]

I'll be honest that this DEFLATE code isn't originally mine (as is already documented) and MS-ZIP is now the only thing that uses it. So if this fixes MS-ZIP, then it's all good by me. I'd 100% prefer to use zlib, but that's a while off.