renderShell paints first and checks for a display target last, and the
stale-data repaint re-requests before that check. requestShellRender runs
`do { renderShell() } while (shellRenderQueued)`, and because a render is
already in progress the re-request just sets shellRenderQueued. So with no
display target the two feed each other: paint, report stale, re-queue, discard
the work, round again, with no yield and no backoff. SHELL_REFRESH_INTERVAL_MS
is 60 seconds, so the timer is not the driver.
Measured on a device with the glasses not presenting, from the app's own
frame-timings export: 40 of 40 recorded frames were render:shell
target=shell-overlay, every one discarded, mean 55 ms and median 53 ms apart,
so about 19 full renders a second all thrown away. Main thread state=Runnable
with utm=22599, and seven ANR records over three days, each "Input dispatching
timed out ... Waited 10001ms". Afterwards the same state produced 5 renders in
100 seconds, and the export showed a normal mixed workload again instead of
one loop crowding everything else out.
Hoisting the check above the paint fixes both halves at once: a frame with
nowhere to go costs nothing, and the stale-data re-request lives downstream of
the early return, so it can no longer feed the loop. The copy of the check that
used to sit after the paint is removed rather than left in place. Nothing
between the two points can change the result, and the compiler agrees: the
early return narrows this.phase, which made the "charging" comparison
unreachable (TS2367).
submitWindowFrame, the sibling path for in-process window frames, already
checks the display target before doing any work.
Not unit-tested. The change is in dashboard-controller.ts, which imports 56
modules, so a test there would mostly exercise mocks, and it is an ordering
defect rather than a predicate, so extracting a shouldPresent() helper would
pass both before and after. The shape that would work is pulling the render
gate and its stale-data retry into a small module taking the display target and
the stale flag and returning whether to paint and whether to re-request.
Open EvenHub once without finishing sign in and the phone side form comes back
on every app start, with no way to close it. Back exits the app rather than
closing the form.
Three things combine:
- EvenHubStoreLayer.paint() opened the credential editor on its first paint
when no credentials were stored. First paint includes the restore of a
still-open EvenHub at startup, so an unfinished sign in put the form back
every time. handleInput already opens it on a click, so the auto open only
removed the choice of when.
- LayerActions.startTextSettingsEdit declares a fifth onCancel parameter, but
the DashboardController wiring forwarded four and the method accepted four.
Nothing failed to compile, because a 4-parameter function is assignable to a
5-parameter function type. That callback resets openCredentialEditor's
phoneEditor guard, so with it dropped the guard latched at "credentials":
the screen timeout hid the panel without resetting the state behind it, and
the form could then be neither closed nor reopened. ios-preview-controller.ts
already forwards all five, so this was the G2 path only.
- The panel has no Cancel control, so the keyboard's Done key on the password
field is the only way to end the edit. On an untouched form that reached
submitCredentials, which called enterLogin("Email and password are
required."), which reopens the editor. A loop with no exit.
The auto open is removed; onCancel is forwarded, stored and fired from
endTextSettingEdit, with finishTextSettingEdit clearing it first so a completed
edit does not also report a cancel; and a wholly empty submission is treated as
a cancel. A half-filled form is a real mistake and still reopens with the
message.
tests/evenhub-signin.test.cjs covers the open and close contract through the
existing load-typescript.cjs loader. Two of its five tests fail against the
current store-layer.ts and pass with this change.
<details>
<summary>Prompts Used</summary>
User: This checkout corresponds to https://github.com/jimrandomh/faceclaw/pull/27. Review this branch.
GPT-6 Astra:
> Found one issue:
>
> - **[P2] Wait for Whisper’s final result before completing dictation** — [FaceclawVoiceController.java:729](/Users/jbabcock/repositories/g2/faceclaw/App_Resources/Android/src/main/java/com/faceclaw/app/FaceclawVoiceController.java:729). Skipping partial decodes leaves short utterances without text until release, but the dialog still finalizes after its existing timeout. If decoding exceeds roughly 2.7 seconds (native stop’s 1.5-second wait plus the dialog’s 1.2-second timeout), auto-send can abandon a short command or send an incomplete longer one. Continue can append the late editing instruction literally instead of refining the message. I reproduced these outcomes with delayed transcript events. These flows should await recognition completion.
>
> Validation: TypeScript and a clean Android build passed. All 21 test failures also occur on `main`. Actual Whisper latency was not tested on hardware.
<hr>
User: Thanks, let's fix that issue.
GPT-6 Astra:
> Fixed. Dictation now waits for native recognition to finish before fallback timers run. Auto-send, Continue, and AI Chat preserve delayed Whisper results; cancellation ignores late transcripts.
>
> TypeScript and Android builds pass. All 11 new regression tests pass. The full suite still has the same 21 pre-existing failures.
</details>
sherpa-onnx's native library already reads an offline Whisper config
(verified against sherpa-onnx/jni/offline-recognizer.cc at v1.13.0);
only the Java wrapper was a package-private stub. This adds a public
OfflineWhisperModelConfig builder and wires base.en (int8) in as a
second on-device model next to Moonshine.
- FaceclawVoiceController: setOnboardModelKind() selects the model
before start("onboard"); defaults to Moonshine. Whisper skips the
700ms live-partial redecode (each call re-encodes the whole utterance
plus ~10s of tail padding) and gates near-silent segments before
decode.
- asr-model.ts: ASR_MODELS registry with per-model download state.
- Settings > Voice: "On-device (Moonshine)", "On-device (Whisper)",
and "OpenAI (Whisper)" for the existing cloud provider (stored
value unchanged); one model row per on-device model.
- Microphones live captions pass "moonshine" explicitly (unchanged).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>