Prepare release

Author: Vincent Wilms

CHANGELOG.md

@@ -1,5 +1,19 @@
 ## v2.1.4 - 2026-07-29
 
+- [fix(writing): emit the group info message](https://github.com/Apollo3zehn/PureHDF/pull/178)
+- [fix(writing): zero the padding of a fixed-length string](https://github.com/Apollo3zehn/PureHDF/pull/177)
+- [fix(writing): key the datatype cache by everything the message depends on](https://github.com/Apollo3zehn/PureHDF/pull/173)
+- [fix(writing): encode datatype names as UTF-8](https://github.com/Apollo3zehn/PureHDF/pull/171)
+
+- [fix(reading): partial stream reads, cache thread safety, and group header re-decode](https://github.com/Apollo3zehn/PureHDF/pull/175)
+- [fix(reading): never decode strings as ASCII](https://github.com/Apollo3zehn/PureHDF/pull/170)
+
+Thanks @Blackclaws for your contributions!
+
+## v2.2.0 - 2026-08-16
+
+## v2.1.4 - 2026-07-29
+
 - [Make tests runnable on Windows and with newer h5dump, with related fixes](https://github.com/Apollo3zehn/PureHDF/pull/160)
   - Add SharedHdf5StateCollection to serialize tests using HDF5/global state, preventing failures from sharing problems.
   - Update xUnit dependencies and add Xunit.SkippableFact for conditional test skipping.

src/PureHDF/VFD/H5StreamDriver.Reading.cs

@@ -94,9 +94,6 @@ private T Read<T>() where T : unmanaged
         var size = Unsafe.SizeOf<T>();
         Span<byte> buffer = stackalloc byte[size];
 
-        // ReadExactly, not Read: Read's return value was discarded, so a stream that delivered fewer
-        // bytes than asked left the rest of the value as whatever the stack slot held. Every length,
-        // address and checksum in the file goes through here.
         _stream.ReadExactly(buffer);
 
         return MemoryMarshal.Cast<byte, T>(buffer)[0];

src/PureHDF/VOL/Native/API.Reading/NativeGroup.cs

@@ -140,16 +140,6 @@ private bool InternalLinkExists(string path, H5LinkAccess linkAccess)
         var segments = isRooted ? path.Split('/').Skip(1).ToArray() : path.Split('/');
         var current = isRooted ? Context.File.Reference : Reference;
 
-        // The first hop of a RELATIVE path is this group, whose object header is already decoded and
-        // held. Dereferencing our own reference to get it - which is what this did - builds a second
-        // NativeGroup and decodes that header again, from the file, on every lookup.
-        //
-        // The cost is proportional to the number of LINKS, not to the depth of the path, because a
-        // group that stores its links compactly keeps one header message per link: measured on a
-        // 1000-link group written by PureHDF's own writer, one LinkExists re-read 30,113 bytes and
-        // allocated 2.1 MB, and 2000 links doubled both. A lookup that missed cost exactly as much as
-        // one that hit, since the whole cost was the re-decode rather than the search.
-        //
         // Only the first iteration can reuse a group we already hold; every later segment names an
         // object not yet resolved, so this is cleared at the end of each pass.
         var group = isRooted ? null : this;
@@ -183,7 +173,8 @@ internal NativeNamedReference InternalGet(string path, H5LinkAccess linkAccess)
         var segments = isRooted ? path.Split('/').Skip(1).ToArray() : path.Split('/');
         var current = isRooted ? Context.File.Reference : Reference;
 
-        // See InternalLinkExists for why the first hop of a relative path reuses this group.
+        // Only the first iteration can reuse a group we already hold; every later segment names an
+        // object not yet resolved, so this is cleared at the end of each pass.
         var group = isRooted ? null : this;
 
         for (int i = 0; i < segments.Length; i++)

src/PureHDF/VOL/Native/Core.Writing/H5NativeWriter.cs

@@ -161,6 +161,8 @@ internal ulong EncodeGroup(
         // No flags, which means the reader applies its own defaults for those thresholds (8 compact /
         // 6 dense in the C library). Declaring values here would state a policy this writer does not
         // implement, since it always stores links compactly.
+        //
+        // See https://github.com/HDFGroup/hdf5/issues/6616 for more info.
         var groupInfoMessage = new GroupInfoMessage(
             Flags: default,
             MaximumCompactValue: default,

src/PureHDF/VOL/Native/FileFormat/Level2/ObjectHeaderMessages/Datatype/DatatypeMessage.Writing.cs

@@ -732,11 +732,8 @@ void encode(object source, IH5WriteStream target)
                 else
                 {
                     using var paddingBufferOwner = MemoryPool<byte>.Shared.Rent(padding);
+                    
                     var paddingBuffer = paddingBufferOwner.Memory.Span[..padding];
-
-                    // Rent does not zero the buffer, so without this the padding written to the file is
-                    // whatever was last in that pooled memory. The stackalloc branch above clears for the
-                    // same reason.
                     paddingBuffer.Clear();
 
                     target.WriteDataset(paddingBuffer);

src/PureHDF/VOL/Native/FileFormat/Level2/ObjectHeaderMessages/Datatype/DatatypePropertyDescriptions.cs

@@ -206,7 +206,6 @@ public override ushort GetEncodeSize(uint typeSize)
 
     public override void Encode(H5DriverBase driver, uint typeSize)
     {
-        // name
         // The specification gives a compound member name no character set field, and the
         // reference library copies it as an opaque NUL-terminated byte string
         // (H5MM_xstrdup in H5Odtype.c), so whatever bytes the caller supplied are stored.

…erMessages/GroupInfo/GroupInfoMessage.cs …es/GroupInfo/GroupInfoMessage.Reading.cssrc/PureHDF/VOL/Native/FileFormat/Level2/ObjectHeaderMessages/GroupInfo/GroupInfoMessage.cs renamed to src/PureHDF/VOL/Native/FileFormat/Level2/ObjectHeaderMessages/GroupInfo/GroupInfoMessage.Reading.cs src/PureHDF/VOL/Native/FileFormat/Level2/ObjectHeaderMessages/GroupInfo/GroupInfoMessage.cs renamed to src/PureHDF/VOL/Native/FileFormat/Level2/ObjectHeaderMessages/GroupInfo/GroupInfoMessage.Reading.cs

@@ -77,7 +77,6 @@ public static LinkInfoMessage Decode(NativeReadContext context)
     {
 
         var (driver, superblock) = context;
-        var b = driver.Position;
 
         // version
         var version = driver.ReadByte();

src/PureHDF/VOL/Native/FileFormat/Level2/ObjectHeaderMessages/Link/LinkInfoMessage.Reading.cs

@@ -12,6 +12,8 @@ namespace PureHDF.Tests.Writing;
 /// every byte >= 0x80 with <c>'?'</c>, destroying the name before any reader is involved.
 /// <para>These assertions go through h5dump rather than PureHDF's own reader, so they check
 /// what actually reached the file.</para>
+/// 
+/// More info: https://github.com/HDFGroup/hdf5/issues/6610
 /// </summary>
 [Collection(SharedHdf5StateCollection.Name)]
 public class NonAsciiNameWritingTests

tests/PureHDF.Tests/Writing/DatatypeTests@NonAsciiNames.cs

@@ -1,4 +1,4 @@
 {
-    "version": "2.1.4",
+    "version": "2.2.0",
     "suffix": ""
 }