A new build's first launch compiles the GPU driver's shaders while the model loads. On a busy machine that ran past the fixed 10 s setup into the measured runs, so listening-at-rest read 1-2.5 s of cpu.work instead of about 10 ms. Setup now waits for session_started in the copy's event log, then 2 s, and at least 10 s. The listening profiles also move the copy's shortcuts to F19 and F18, as the dictation profiles do.
docs/perf.md and site/README.md: authors write their own text; strobe report experiment <exp> --format pr --update-pr <n> places strobe's generated block. Window cases expect the Hex title and their pane's heading, the site bench its page title and address, so a block on another screen is refused WrongScreen.
strobe 0.6.0 keeps each variant's compiled shaders across its launches by default (gpuCache warm). docs/perf.md says what that means for the cases and when a target would need gpuCache cold.
* Accessibility: name Hex's panes and controls for screen readers and strobe
* Rune c35fe19; pane_header beside title_bar; rows report selected only when selected
* Drop the unused title_bar import
* Tests: microphone row at the tall size; say what the tree held when a key is missing
* Tests: read the microphone field with its menu closed
* Rune f6799d3
* Windows listener: stop polling every 5 ms while it records
With the microphone streaming, its chunks (about every 10 ms) wake the loop while it records too, as they already did while idle: shortcut events and commands are taken at the next chunk, and a release's tail ends on one. A shortcut's audio boundaries come from its event's own time, so nothing about the clip changes. Only an opening or recovering microphone is still polled. Adds perf/dictation.bench.ts (recording) on HEX_AUDIO_FIXTURE.
* bench: name the fixture PR
On Windows a prewarmed overlay keeps no pixels while hidden, only its
surface. The picker's surface covers the whole primary display, so a tray
Hex held it in GPU memory from launch. The picker is now made the first
time it opens. It shows only before the capture it starts, so that
capture still knows its window. The idle memory case also reports
gpu.vram.
Import the bench SDK from @hona/strobe (the old @hona/strobe-bench name is deprecated). The idle head reports canvas.frames and canvas.interval.p95, the show's own WebGPU frames at its own rate, and the load case gpu.compile. Vite emits hidden source maps when strobe builds (STROBE_SOURCEMAP=hidden), so profiles name source lines; deploy builds are unchanged. Docs: any commit works as --base now (strobe overlays head's harness), report --format pr for PR tables, and large seeds are shared read only.
Tray targets use strobe's windowless native target; a hidden launch ends on the local API's start line; listening cases with released and open microphone profiles.
* Let strobe launch an isolated Hex
Scope the Windows instance event to a custom data root, read ~/.config/hex from USERPROFILE on Windows, add HEX_DISABLE_UPDATES, name the UI and remaining long-lived threads, and add the strobe profile and feature, strobe.json, perf benches, and docs/perf.md.
* strobe.json: keep the copy out of the Hugging Face cache
* strobe.json: keep seeded history
* Windows instance event: keep the default event for the default root
Naming the default data root in HEX_APPLICATION_SUPPORT_DIR must not let two apps share it.