windows-v0.0.10

FluidVoice Windows 0.0.10 beta release.
- Fixed the memory growth behind multi-day slowdowns (#917). The speech engine's CUDA backend retained a few MB of GPU-driver memory on every dictation, which built into gigabytes over days of uptime and made the recording overlay take seconds to appear. The engine is patched (fix maintained in our parakeet.cpp fork; upstream PR to follow) and the corrected CUDA runtime downloads automatically on next launch. Dictation on NVIDIA GPUs is also ~12-15% faster with the fix.
- As a safety net, FluidVoice now restarts itself when idle if its memory has grown past a threshold, or after a day of uptime. The restart is invisible: it never happens while dictating or with the window open, and the replacement warms its models before you return.
- Long-running sessions now rotate the log file instead of growing it without bound.
- Redesigned Settings and the Insights page, with a greeting that uses your Windows name.
- Fixed dictations silently failing to paste. Clipboard managers such as Windows Clipboard History could lock the clipboard the moment FluidVoice wrote to it, so the paste did nothing and the transcript was then lost. FluidVoice now marks the write so those tools skip it, checks that the text actually landed, retries once when it did not, and leaves the transcript on the clipboard instead of overwriting it.
- Crashes are now reported on the next launch instead of disappearing silently, including cases where the app is killed outright.
- When Fluid Intelligence fails, the reason is now recorded so the failure can be diagnosed. Still anonymous, still opt-out, and no extra events are sent.