Version 1.0.5.0

What's Changed

• Stop flagging every other token after a think block as thinking by @MBrekhof in #11
• Dequantize each weight column once per prefill, not once per row by @MBrekhof in #10
• Store the LM head verbatim [vocab, in]; stop transposing it on load and mis-sizing it from our own exports by @MBrekhof in #14
• Await the producer when a data-loading enumeration is abandoned early by @MBrekhof in #16
• Keep block activations alive until the block context is disposed by @MBrekhof in #17
• IAcceleratorPlugin: a plugin seam for accelerator-backed training by @MBrekhof in #18
• Sanitise every block scale in the quantization agreement fixture by @MBrekhof in #19
• Publish the activation factory's last-hit slot as one atomic pair by @MBrekhof in #20
• Remove the per-column GPU inference path by @MBrekhof in #21
• Block the F32 matmul over rows, the way the F16 prefill kernel already is by @MBrekhof in #22
• Unpack Q4_K, Q6_K and Q5_0 weights with vector ops, not through a stack buffer by @MBrekhof in #24
• Widen the quantized prefill microkernel to a 4x3 output tile by @MBrekhof in #25
• Wire LoRA into the training path: per-layer adapters, frozen base, merge for export by @MBrekhof in #26
• Route the training LM head and trainable linear backward through the matmul kernel by @MBrekhof in #27
• Ship the GPU training engine as an accelerator plugin, with attention on cuBLAS by @MBrekhof in #28
• Stop SharpMindChatClient leaking tool-call markup and dropping returned calls by @MBrekhof in #29
• Default ToolsFolder instead of resolving PluginsFolder twice by @MBrekhof in #30
• Fix OpenAI server answering a prompt the caller never sent by @MBrekhof in #32
• Cover BuildCompactToolList's compaction path with a regression test by @MBrekhof in #33
• Make the server's startup preload actually run by @MBrekhof in #34
• Warm the matmul kernels during the server's startup preload (1.67x on the first request) by @MBrekhof in #35

Full Changelog: v1.0.4.0...v1.0.5.0