Release v0.6.0 — macOS (Apple Silicon) port · Fangyuan025/Chaty

Chaty now runs natively on Apple Silicon Macs, alongside Windows.

Metal GPU backend — selected per target via a feature-multiplexer crate; Windows keeps Vulkan unchanged. On unified memory, all layers are offloaded when the model fits (recommendedMaxWorkingSetSize budget), with P-core-only worker threads.
Native window chrome — traffic lights (titleBarStyle: Overlay), Dock-icon reopen, menu-bar tray, Cmd+Shift+Space global hotkey.
Clean quit on every path — tray Quit, app-menu Quit and Cmd+Q no longer trip ggml/ONNX teardown crashes ("Chaty quit unexpectedly").
.dmg packaging with entitlements (mic, JIT, library validation for the bundled ONNX dylibs); CI builds the dmg headlessly via hdiutil.
Native microphone capture (CoreAudio/cpal) — WKWebView never exposes capture devices to embedded apps, so recording bypasses it entirely; devices are scanned (no phantom default-device failures) and the system mic consent is requested properly.

Gemma 4 — native renderer for the <|turn>role …<turn|> format, <|think|> thinking control, <|channel>thought reasoning folded into the think panel, turn-boundary stop insurance.
Qwen 3.5 / 3.6 — pre-opened <think> handling (synthetic open tag for the UI), reliable no-think (empty think block), detected by GGUF architecture so community finetunes with custom templates behave too; the dead /no_think soft switch is never sent to 3.5+.
Robust template fallback chain — embedded template → system-role folding → per-architecture built-in → ChatML, so unusual GGUFs still chat.
Families covered: Llama 3, Gemma 3, Gemma 4, Qwen 3, Qwen 3.5/3.6.

Synchronous eject before load — switching models fully tears down (and verifies release of) the old model before the new one loads; no more unified-memory swap freezes. mmap is disabled on macOS (Metal-wired pages of mmap'd MoE models never returned to the kernel).
Model load progress bar (eject → weights % → ready) in the titlebar and model chip.
Pre-flight guard — models that cannot physically fit in RAM are refused with a clear message.
Context auto-fit — "Auto" now uses as much of the model's trained context as memory allows (KV-cache-size aware); custom values are capped to fit, with a visible notice when clamped; the settings slider adapts to the loaded model's trained length.

Stop reason shown after each reply (finished / length / context full / stop sequence / cancelled).
Focused thinking view — while reasoning streams, a small window follows the newest text with older lines fading out; expandable as before.
Circular context-usage ring (amber > 80 %, red > 95 %).
Unlimited reply length by default (opt-in cap in settings); "Reload to apply" button under the context setting.
Playable HTML preview — single-file web games work (keyboard focus + localStorage shim in the sandbox).
Mermaid: theme-aware, looser parsing, and visible error messages instead of silently showing raw code.
Higher-contrast light theme; "Open models folder" in the model menu; backend errors are bilingual (中文/English).
GPU/memory usage in the hardware panel now reports the app's real footprint.

Cross-platform release CI — pushing a vx.y.z tag builds the Windows installer and macOS dmg and publishes both to one GitHub Release (version consistency is checked against the tag).
scripts/bump-version.sh syncs the version across package.json, tauri.conf.json, Cargo.toml and Cargo.lock.
The in-app updater picks the right asset per platform (.exe / .dmg).

Provide feedback