voxtype-bin: update to 1.0.1

[spencerbull]

Updates voxtype-bin from 1.0.0-2 to 1.0.1-1 (upstream release).

Packaging changes

• pkgver=1.0.1, pkgrel=1; all binary checksums taken from the release's SHA256SUMS.txt, whose signature I verified against the Voxtype CI signing key (9CCF7915B750CAE8B095ED1AA3FC9F33FD209279) already listed in validpgpkeys. Support files fetched from the v1.0.1 tag and hashed locally. Full makepkg --verifysource passes (all 50 sources: checksums + GPG signatures).
• Fixes a latent 1.0.0 packaging bug: quickshell/voxtype-shared/qmldir has registered StyleLoader and RecipeRenderer since v1.0.0, but the package never shipped those two QML files, so import voxtype-shared 1.0 referenced missing components. They're now in source= and installed with the rest of the module.
• .install post-upgrade notes rewritten for 1.0.1; dropped the 1.0.0 known-issues block (the Quickshell pointer-capture issue #440 is closed upstream).
• The CUDA/MIGraphX ONNX Runtime provider .so companions are byte-identical to 1.0.0 (same ORT 1.24.4 redistribution), which matches expectations for a patch release.

Not changed: the voxtype-onnx-rocm compat symlink (PKGBUILD comment says "drop in v0.8.0") is still shipped — dropping it is user-visible and worth its own decision, not a ride-along in a version bump.

Upstream changelog (v1.0.1)

Voxtype 1.0.1 fixes what people reported in the week after 1.0.0. Four of these change what the audio pipeline actually hands the model, so this one is worth taking even if none of the bugs bit you.

The resampler was aliasing — resampling used plain linear interpolation with no anti-aliasing filter, folding everything above 8kHz back into the speech band on 48kHz/44.1kHz mics; the fractional read position also reset ~100×/sec, and up to 21ms was dropped off the end of every recording. All three fixed, nothing to configure. Verified upstream with byte-identical transcripts across 16/44.1/48kHz.

Media ducking fades — ducking ramps instead of snapping, without holding up the microphone. New [audio] duck_media_fade_ms (0 restores instant behavior).

Model downloads resume — interrupted downloads resume where they stopped, stalled ones abort, and a partial file never lands at the final path.

Push-to-talk stays unstuck — keyboard hotplug mid-press no longer leaves recording running (#556), the dotool fd leak that spun a CPU core at 100% is fixed (#445), and dropped evdev events no longer desynchronize pressed-key state.

Idle model eviction runs now — the eviction timer sat on a tokio::select! arm that could never win, so the feature had never executed. A daemon left running all day now actually releases model memory when idle (#644).

Bug fixes — audio stream errors reported instead of silently truncating recordings (#642); remote mode uses remote_model (#496) and honors language = "auto" (#619); voxtype configure honors -c/--config (#595); configure TUI no longer freezes on ALSA dsnoop devices (#541) or misreports installed sensevoice/moonshine models (#662); filler filtering no longer eats "um" in Portuguese/German (#566); zero-delay eitype output stops truncating on KDE (#552); eager processing stops using Whisper for chunks when a non-Whisper engine uses on-demand loading (#618); setup variant --to refuses binaries that can't run on the host CPU (#633); setup model --set routes by engine instead of crash-looping the daemon (#610); aarch64 OSD binaries now ship (#579).