Fix five regressions on LTRData.DiscUtils-initial (registry, NTFS, dynamic-disk hang, sparse-stream)

[devedse]

Fixes five regressions currently on LTRData.DiscUtils-initial. With all five, the full LibraryTests suite passes (581 passed, 1 skipped, 0 failed, no hangs). Without them the suite cannot build/create a hive (#2), fails NTFS/ExFat reads (#3, #5), and hangs indefinitely (#4).

1. SubKeyIndirectListCell — InvalidCastException on mixed ri/li lists

LinkSubKey/UnlinkSubKey branched on ListType and hard-cast every entry, but a list can contain a mix of sublists and key nodes (the read paths already tolerate this). Deleting/adding a subkey under such a list threw InvalidCastException: KeyNodeCell -> ListCell. Fixed by dispatching per entry on the actual cell type, mirroring the read paths.

2. Bin.AllocateCell — 8-byte alignment check regressed to mod 7

Cells are 8-byte aligned, so the check must be (size & 0x7) != 0. It had been changed to (size % 0x7) != 0 (mod 7), rejecting ~6 of every 7 valid sizes including the ~80-byte root cell, so RegistryHive.Create() itself throws Invalid cell size. Restored to & 0x7. (Introducing commit in a comment below.)

3. NtfsFileSystem.GetFileLength — FileLengthFromDirectoryEntries option ignored

The default-data-stream shortcut returning the directory entry's RealSize was removed, making the option a no-op and breaking NtfsFileSystemTest.GetFileLength. Restored the shortcut. (Introducing commit in a comment below.)

4. DynamicStream — infinite loop in sector scan (the hang)

FindNextPresentSector/FindNextAbsentSector advance with pos += (8 - sectorInBlock & 0x7) * Sizes.Sector. By C# precedence that is (8 - sectorInBlock) & 0x7, not 8 - (sectorInBlock & 0x7): at a byte-aligned, empty bitmap byte the original 8 - sectorInBlock % 8 advanced by 8 sectors, but the rewrite yields 8 & 7 = 0, so pos never advances and extent enumeration / the dynamic VHD+VMDK builders loop forever. Parenthesized to restore the semantics. (Introducing commit in a comment below.)

5. SynchronizedSparseStream — Position never advances

The synchronized wrapper set content.Position = Position before each op but never wrote the advanced position back, so Position stayed frozen and reads/writes through the wrapper re-processed the same offset (e.g. ExFat...ReadLongSparseLimited). Position now delegates to content.Position under the lock, fixing every read/write path at once. (Introducing commit in a comment below.)

[devedse]

Note on fix #2 (Bin.AllocateCell): the (size % 0x7) regression was introduced in commit f7bf755 ("Minor code cleanups and optimizations", 2026-06-05), which changed size % 8 != 0 → (size % 0x7) != 0.

That is also where CI on LTRData.DiscUtils-initial started failing — the preceding commit ("Minor optimizations", 2026-06-02) was green (3/3), and f7bf755 went red (2/3). Because it makes RegistryHive.Create() throw on the root cell, it breaks every registry write on the current default branch, not just this PR.

[Olof-Lagerkvist]

Oh that was a good find! I have not noticed that CI fails now because I did not get any notifications about it. I get notifications about it from other repositories though. Weird. I need to check what is happening here!

[devedse]

Yeah, my AI found it 😄

Also have a good review of the other fix I did as it's also AI authored and I don't really have the knowledge on this part to see if it makes sense.

[devedse]

Hmm it does seem the build is still failing even after this fix. I'll see if I can figure out what's going on.

[Olof-Lagerkvist]

Yes, the bugs were mostly introduced by some AI optimizations. Though, they feel in most cases safe as long as CI tests still complete successfully. Which I thought they did. 😅

But I think I also need to add CI tests for Linux. There have been several cases lately where tests work in Windows but fall when I run them in Linux and need to address something. Usually just the rest itself, but sometimes it reveals issues deeper in the library.

[devedse]

Note on fix #3 (NTFS GetFileLength): the regression came from commit a5516351132e259b5149030e000f4789f78894d2 ("Bugfixes for FileExist and GetFileLength"), which removed this shortcut from GetFileLength:

if (NtfsOptions.FileLengthFromDirectoryEntries && attributeName == null && attributeType == AttributeType.Data)
{
    return (long)dirEntry.Details.RealSize;
}

That made the FileLengthFromDirectoryEntries option a no-op. The FileExists change in the same commit is correct and is left intact here; only the GetFileLength shortcut is restored.

[devedse]

Introducing commits for fixes #4 and #5:

• Only use Path.Combine on netcore #4 (DynamicStream infinite loop): f7bf7550ff632be3169e0515f3961206667df14e — "Minor code cleanups and optimizations" (the same commit as fix Read only up to MaxSize for VMDK header #2). It rewrote 8 - sectorInBlock % 8 to 8 - sectorInBlock & 0x7; the missing parentheses change the value to 0 at byte boundaries, so pos never advances.

• VHDX LogEntry.TryRead() Bug Fix and Improvements #5 (SynchronizedSparseStream.Position): 4b5b75df — "Synchronized wrappers for disks and streams", which added the wrapper but never advanced Position after a read/write.

For #4 specifically, note the 7 - x & 0x7 mask forms in the same files are coincidentally fine ((7-x)&7 == 7-(x%8) for all x), but the 8 - x forms are not (8-(x%8) ranges 1..8 while (8-x)&7 ranges 0..7), which is why only the two 8 - sites needed the fix.

[devedse]

The build now succeeds so it all seems to work 😄

[Olof-Lagerkvist]

I merged the commits with some changes to avoid reintroducing some bugs that were fixed previously. I also added timeouts to CI tests to avoid this scenario in the future where new bugs that hang unit tests go unnoticed because they never finish and never report failure.

Closing this PR.

Thanks a lot for finding and solving these bugs!