* Hardened file parsers against malformed input: SoundFont files with chunk sizes past the end of the file, or out-of-range/out-of-order preset, instrument, bag, generator or sample indices now throw `InvalidDataException` instead of attempting multi-GB allocations or throwing `IndexOutOfRangeException`/`OverflowException`/`ArgumentException`; WAV cue chunks with a corrupt cue-point count no longer over-allocate or throw; an MP3 whose first frame is too small for the Xing/Info fields it flags is treated as having no Xing header instead of throwing from the `Mp3FileReader` constructor (#1502)
* `SoundFontSampler` skips SoundFont regions with a zero (or out-of-range) sample rate instead of throwing on every note-on (#1502)
* `MidiFile` now pairs note offs with note ons in constant time, so files with many held notes no longer load in quadratic time (#1502)
* `WaveInBuffer`, `WaveOutBuffer` and `MidiIn` no longer free buffers the driver still owns when unprepare reports they are still in use, and `MidiIn` bounds sysex reads by the buffer length; `Vst3Module` loads plug-in binaries by full path (#1502)
* **Breaking:** SFZ `#include` and `sample=` paths that are absolute or UNC/network paths (e.g. `\\server\share\x.wav`, which on Windows can leak credentials) are now refused; relative paths, including `..\Samples\...`, still work. `FileSfzIncludeResolver` and `FileSfzSampleLoader` gain a `rootDirectory` constructor argument to sandbox untrusted SFZ files to a directory, and the shared rules are exposed as `SfzFilePath.Resolve`. `SfzParser` now limits `#include` count and `#define`/`#include` expansion size, and include files are capped at 16 MB (#1501)
* Fixed native buffer overruns and use-after-free bugs: `WasapiOut` could copy past the WASAPI render buffer if an `IWaveProvider` over-reported its read count; DMO `MediaBuffer.Length` accepted values larger than the buffer; `MediaObject.GetInputType`/`GetOutputType`/`GetOutputCurrentType` returned media types pointing at freed memory (`GetWaveFormat()` on them now throws `InvalidOperationException`); `Vst3Plugin` overflowed its native buffers for a huge `maxBlockSize`; macOS `CoreAudioPlayer` could write past the HAL buffer list when the device's stream count shrank; `MediaFoundationReader` could `PropVariantClear` uninitialised memory for media without a duration (#1501)
* **macOS wrappers (preview)**: Fixed Extended Audio File Services reader/writer settings contract to be more accurate after the nullable change landed in #1486. New APIs in the `AudioFileLibraryInformation` class were also added. (#1500)
* **Breaking:** `WaveFormat.FromFormatChunk` (and so `WaveFileReader.WaveFormat`, `AudioFileReader` and anything else reading a WAV fmt chunk) now returns the most specific format type instead of always returning a `WaveFormatExtraData` — a `WaveFormatExtensible`, `AdpcmWaveFormat`, `Gsm610WaveFormat` or `Mp3WaveFormat` where the encoding and extra data allow, a plain `WaveFormat` where there is no extra data, and a `WaveFormatExtraData` only for an encoding with no dedicated subclass. This is the same choice `WaveFormat.MarshalFromPtr` makes, so a format read from a file and the same format read from a native pointer now come back as the same type. Code casting a reader's `WaveFormat` to `WaveFormatExtraData`, or unpacking a `SubFormat` GUID out of its `ExtraData`, should pattern-match on `WaveFormatExtensible` (or call `AsStandardWaveFormat()`) instead (#1497)
* `WaveFormat.MarshalFromPtr` now decodes a WAVE_FORMAT_MPEGLAYER3 block into an `Mp3WaveFormat` rather than a `WaveFormatExtraData` when its cbSize accounts for the 12 MPEGLAYER3WAVEFORMAT bytes (#1497)
* A `WaveFormat` subclass decoded from a format block now reports the `ExtraSize` its `Serialize` actually writes, so a block whose cbSize over-reported the extension (drivers pad it, and a fmt chunk can be padded too) no longer round-trips into a fmt chunk shorter than the cbSize it advertises (#1497)
* Added `BufferedWaveProvider.CompleteAdding()` and `IsAddingCompleted`, so a producer can end the stream: `Read` drains the buffer, returns one unpadded short read even with `ReadFully`, then 0. Added `UnderrunSilenceBytes`, `DiscardedBytes` and `ResetCounters()` to show silence inserted on underrun and audio dropped on overflow (#1492)
* Added `MixingSampleProvider.CompleteAdding()` and `IsAddingCompleted`, so a `ReadFully` mixer ends once its current inputs have ended; `AddMixerInput` after it throws `InvalidOperationException` (#1492)
* **`WdlResamplingSampleProvider` now defaults to sinc resampling** - a new `WdlResamplerQuality` constructor argument selects the algorithm, and the default is now `Sinc64` (64-tap windowed sinc) instead of linear interpolation with IIR filtering, which lost about 2 dB at 10 kHz and 20 dB at 18 kHz and let aliasing through. This is a behaviour change: output samples differ, CPU use is a little higher and about 0.8 ms of latency is added. Pass `WdlResamplerQuality.Linear` to keep the old output exactly. See [Resampling](Docs/Resampling.md) (#1491)
* Added `WdlResamplingSampleProvider.HoldBackOnShortReads` for sources that return short reads mid-stream (e.g. `BufferedWaveProvider` with `ReadFully` off): in sinc mode it holds back the last few output samples of a short read until their input arrives, instead of computing them from silence. `WdlResampler.ResampleOut` gains a matching overload with a `flush` argument (#1491)
* `NAudio.Asio` is now Native AOT compatible. The driver binding no longer uses reflection or runtime marshalling: the IASIO vtable is bound with typed function pointers, and channel info, clock sources and callbacks go through blittable structs. The public API is unchanged. A new test drives the binding against an in-process fake driver, and CI runs it as a Native AOT executable. Also fixed `AsioDriver.DisposeBuffers` freeing its callback table twice when called again, e.g. by `ReleaseDriver` after a failed `AsioDevice.Reinitialize` (#1487)
* **New package: NAudio.Nam** - native C# inference for [Neural Amp Modeler](https://www.neuralampmodeler.com/) captures. Loads A2 (A2-Lite and A2-Full), A1 WaveNet and LSTM `.nam` models and runs them as an allocation-free `NamEffect` (zero latency at the model's sample rate, with built-in sinc resampling at other rates), or through `NamProcessor` for ASIO callbacks. Output matches NeuralAmpModelerCore to within float rounding. See `Docs/NeuralAmpModeler.md` (#1498)
* Added `CabinetEffect`, a zero-latency guitar/bass cabinet simulator: it loads an impulse response from a WAV file at any sample rate, resamples it to the stream rate, and optionally trims its silent tail and normalises its level (#1498)
* Added a zero-latency mode to `PartitionedConvolver` (new constructor overload) and `ConvolutionReverbEffect` (`ZeroLatency`), which applies the first partition of the impulse response directly so there is no partition delay (#1498)
* `PartitionedConvolver` (and so `ConvolutionReverbEffect`) is about 2-3x faster: the frequency-domain multiply-accumulate is now vectorised (#1498)
* Fixed `WdlResampler` (and so `WdlResamplingSampleProvider`) allocating a new input buffer on most calls when the block size varied; the buffer now only grows, so steady-state resampling is allocation-free in both output-driven and input-driven mode (#1498)