Writing a fractional alpha onto the overlay record for every paint item
boxed a heap number per item per frame. The allocation gate caught it:
the generic 60-entry workload went from 2,838 to 12,614 B/frame. The
record now keeps only the frame stamp and a pointer to its paint item,
and isOverlayEntryQueued multiplies the alphas when it is asked. The
gate is back to main's numbers on Node 24.
With Vessels and Detection on, a vessel with an ambient card on the map
also got its detection callout next to the bracket. The callout lane
paints beneath the cards, so its text showed through the card's plate
and covered the speed and course.
The callout now gives way only to a card the shared overlay actually
paints in the same frame. The overlay paints part of the published
cards (shared budget, collisions, UI chrome), so a vessel whose card
was dropped keeps its callout.
The detection lane paints before the cards publish their paint rects,
so getOverlayPaintRect has nothing for the current frame there. The new
isOverlayEntryQueued answers from the paint queue that the overlay
solves before painting. Vessel detection objects name their card in
cardSourceId/cardEntryId.
Arabic catalog strings carry optional marks ("تتبّع"), but people type
without them, so "تتبع" found no suggestion. Fold the marks out of both
the draft and the candidates before comparing.
The quick commands "Reset to globe view" and "Show earthquakes" reached t()
through a .map(), so the catalog check never saw them and every language
showed them in English. They are now marked with msg() and translated.
Review of the 71 strings from #1507 in all eleven languages:
- zh-Hant is regenerated from zh-CN with OpenCC s2tw, as its guide says;
40 entries had Taiwan vocabulary instead.
- ar "Dock to the right" names the left side, where RTL puts that rail.
- ja "What am I looking at?" no longer asks what the assistant sees.
- Provider Settings, dock, provider, live and map use each glossary's terms.
- de drops the du-form; ja drops spaces between Japanese and Latin text.
- Calques, register and the example commands read naturally.
The heatmap now takes over above 4,500 km instead of 3,000 km, with
the cross-fade between 3,600 and 5,400 km. The markers band caps at
40,000 detections, which covers the ~40,000 in the padded view over
Africa at the low end of the fade, so weak fires still stay as markers.
To keep that rebuild near its old cost, markers outside the close band
read the anchor already built for each fire instead of the ground floor
cache: only the close band warms floors, so the lookup there always
missed. A 40,000-marker rebuild now costs ~55 ms, down from ~120 ms.
The 750–3,000 km markers band now draws up to 20,000 detections instead
of 2,500. At the 2,400 km end of the cross-fade the padded view over
Africa holds ~21,000 fires, and the old cap kept only those above
~17 MW, so most heat spots faded into empty ground as the heatmap went
out. A rebuild of that many costs ~60 ms, and only in the densest views.
The far-zoom heatmap now uses the ½× grid (4096 × 2048, ≈10 km per
pixel, one-pixel blur), so a lone fire's spot matches its marker at the
3,000 km hand-off. The spot size slider and its share option are gone.
Markers and heatmap are both kept built at every zoom and cross-fade by
camera height between 2,400 and 3,600 km on every frame, so a fast zoom
never shows a beat with neither. The heatmap fades by imagery alpha:
hiding a Cesium imagery layer unmaps it from the globe tiles, and showing
it again re-mapped them late, which was the gap at the band edge. Under
the markers the heatmap is built 150 ms after the data or filter changes.
A ½×–4× slider (log2 axis, 5 % steps, default 1×) under the intensity
filter sets how large far-zoom heat spots are. It scales the grid, not
the blur: 2× is a 1024 × 512 grid with the same one-pixel blur, so each
spot grows exactly and weak lone fires don't fade as they widen. A change
rebuilds only the heatmap. Share links keep it as w.s (percent, 50–400).
The layer row now takes a list of sliders.
One provenanceLine() builds STATE | text | LAYERS for the HUD and
get_hud_caption. hasWord is one lookaround regex, capAtWord one grapheme
pass. Feed-state helpers are imported from layerSnapshot.js directly, and the
kept-whole sanitizer cases are one table.
Dropping each detection whole into its nearest pixel snapped spots by up
to half a pixel and left weak lone fires as squares or plus signs once
thresholded. Share each detection's weight between the four nearest pixel
centres instead (linear binning, Wand 1994): spots sit at the true
position and blur round, at the same build cost.
A dial click shorter than the device start-up let suspend() run on a still-suspended context, so the later resume() left it running with no nodes. Cover paused, connecting and voice-fallback releases in radioNextState tests.
At 10 km per pixel each fire was a marker-sized spot that faded into the
imagery from the globe view. A 2048 px grid with the same one-pixel blur
makes every spot exactly twice as wide (~60 km for a lone weak fire),
still separate, and the build is three times faster.
The heat raster was 2048 px wide with a 3 px blur, so every fire spread
over ~200 km and the far-zoom map read as clouds next to the markers.
Double the grid to ~10 km per pixel and blur by one pixel: a lone fire
is now a spot about as wide as its marker at the 3,000 km hand-off.
A release that only selects a station now stops the looping noise source and suspends its AudioContext. The qa-radio dial scenarios start from an active stream, or expect a silent release from a stopped Radio.
Han and kana have no word edges, so a CJK forbidden word now counts anywhere
in the reply. The proxy no longer counts words: the model reads Rue de Rivoli
as one, so only the 80-character word-boundary cap remains.
Also restore src/voice/gevActions.test.mjs, reformatted by mistake in 818d2c82.
The heat scale is set by the hottest regions, so most of the world sits
at the low end. The ember edge was a dark red at 35% opacity and nearly
vanished into forest and savanna imagery: over the Americas the spots
were there but hard to see. The edge is now a bright red starting at
58% opacity, which reads at least as well as the old blue did.
About 90% of detections are under 20 MW, so the old blue and teal
stops painted most of the map in cold colors that do not read as fire.
The ramp now runs ember, red, flame, orange, gold and white-hot at the
same thresholds, each stop lighter than the one below, so the brightest
marks are the strongest fires. The heatmap uses the same colors.
- The proxy keeps combining marks, in-word apostrophes and name dots
(ar vowel signs, fr d'Orléans, ja ・), normalizes to NFC, caps every
summary in characters and cuts at a word boundary.
- feedState is nullable instead of a `none` token, and the summary has at
most five words: "exactly five" made the model pad short views with
"no data" or "unavailable".
- The HUD caches only the model's words and adds the current warning on
every paint; an RTL locale gets a leading right-to-left mark.
- The echo check also rejects an all-Latin reply in a non-Latin language,
translated layer ids and, with every feed healthy, the UNAVAILABLE word;
places are matched by word and one-word layer names are skipped in
Latin-script locales.
- The warning tag names the worst layers first; a partial analyst card
shows one partial chip.
Releasing the tuning dial always started the chosen station, plus tuning
static, even when nothing was playing. It now follows the Previous/Next
rule: the release plays only if a stream was loading, buffering or playing
when the drag began; otherwise it just selects the station.
The privacy note no longer lists finishing tuning as a way to connect and
now names voice instead, in all locales. Docs, changelog and comments are
tightened to match.
Main moved the layer rows to Vue, so the fire confidence switch and
minimum-FRP slider now render through RowSwitch and RowSlider on the
kit's UiSegmented and UiSlider, with the same classes as before. The
fire filter strings are translated into Arabic.
main moved the Radio markup into Vue views and rebuilt the UI tests on
stores. Resolved:
- src/ui/templates/context.html: deleted in main; the new privacy note now
lives in src/ui/views/panels/RadioPanel.vue.
- src/ui/radioControls.test.mjs: main's store fixture; the Next/Previous
autoplay test drives store.act('cycleRadio').
- src/voice/gevRealtime.test.mjs: main's createTestVoiceUi; the stop and
select handoff cases stay one table-driven test.
- src/data/radioNextState.test.mjs reads the published gev-radio view.
- src/i18n/locales/ar.json: the new note translated for the Arabic locale.
- _relocalize() now drops the received reply and aborts one in flight, so
the next tick asks in the new language instead of keeping old-language
text under the new-language state tag.
- The English-echo check also allows state words inside a place label
(Estación Loading Dock).
- Reword the non-English instruction so it reads for every language name.
- Add the two new test files to the format scope.
- Skip the model when the view has layers but no place, street or nearby
label: with layers alone it invented a place.
- The English-echo check treats only Latin letters as part of a word, so
ids flush against kana or Han and hyphenated names are caught; it skips
nominal, state words a locale keeps in English and names inside place
labels.
- Outside English, per-layer state ids are left out of the request.
- The ja/zh cap steps back to the start of a Latin word, not the last space.
- Fold the CJK cap cases into hudSummaryResponse.test.mjs.
Restore the two fixes the previous commit moved out:
- A tuner release onto a station the directory dropped mid-drag stops the
stream the drag paused, publishing one state with the gesture's origin,
so the button reads PLAY and a voice handoff is cancelled.
- The transport buttons offer Pause while a stream connects, matching
what a click does.
Their tests now use the shared Radio layer harness.
- Selections take the playback fallback as a `playFallback` option and
set it after releasing a stale stream, instead of callers arming state
that the release then had to save and restore.
- Move the tuner `station-unavailable` release and the Pause label while
a stream connects out of this branch: both were wrong on main already
and go in a separate change.
- Share the Radio layer test setup in src/testSupport/radioLayerHarness.mjs
and fold overlapping Next/Previous tests (13 to 7).
- One-line comments and a two-line changelog entry.
Second follow-up review fixes:
- Japanese and Chinese summaries are capped at 20 characters only, without
the five-token cut first, and the cap steps back to a space instead of
splitting a Latin word.
- The HUD request sends localized layer names in feedProvenance too and
leaves the English narration note out outside English. A non-English reply
that still echoes a feed-state id or a translated English layer name falls
back to the deterministic line.
- A camera move during typing shows the deterministic line, not a fragment.
- A view with no place labels and no enabled layer does not ask the model.
- Behaviour tests on IntelHUD replace the regex checks of hud.js; docs and
the instruction wording updated.
Follow-up review fixes:
- The HUD line is STATE | text | LAYERS for the AI summary and the
fallback line, so the narrow minimal and cyber layouts show both the
warning and the summary. get_hud_caption builds its caption the same
way. The unused "{state} {layers}" key is pruned from the catalogs.
- The 20-character cap applies only when the request locale is Japanese
or Chinese; other languages keep five words around a Han place name.
- enabledLayers names are localized too, and the instructions name layers
only from enabledLayerLabels, never from feedProvenance or its note.
- The keyless test restores GEV_RATELIMIT_OPENAI_PER_MIN again; comments
and docs updated.
main added de, fr, it, ja and pt with the old "after you press play"
note; give them the updated wording. main also removed the compact
Context Radio player, so the Next/Previous tests stop checking its
button.
Review follow-ups for HAL-722. A Next or Previous that only selects a
station now leaves nothing of the previous stream behind:
- Selecting another station without playing stops a paused or failed
stream, so the panel, Stop, voice summaries and the audio element agree
on "stopped". A fallback the caller armed for the new selection is kept.
- A silent UI step arms no broadcaster fallback, and a user Play of the
selected station drops one left by a cancelled voice step. Voice steps
pass `fallback: true` because voice starts the prepared playback later.
- A user step to another station emits a `select` playback control, and
the realtime voice handoff treats it like Pause and Stop.
- A tuner release onto a station the directory no longer has stops the
stream the drag paused, publishing one state with the gesture's origin,
so the button reads PLAY and a voice handoff is cancelled.
- The transport buttons offer Pause while a stream connects, matching
what a click does.
- The privacy note names every action that connects to a broadcaster
(play, finishing tuning, choosing a station on the globe) in all
locales.
The Next/Previous layer tests live in src/data/radioNextState.test.mjs.
Review fixes for the HUD summary language change:
- The translated feed-state tag now leads the one-line summary and the
fallback line, so the readout's ellipsis cannot hide it.
- get_hud_caption checks the reply's feedState, returns it and appends the
state tag; a reply naming another state keeps only the place.
- The model gets the layer names the user sees (localizeLayerName).
- Japanese and Chinese summaries are capped at 20 characters; pt, zh-CN and
zh-Hant get explicit language names.
- A partial view is narrated as incomplete coverage, not as not live.
- provenanceTag and hudSummaryMatchesProvenance move to layerSnapshot.js so
the tools package can use them without browser globals.
- docs/CURRENT-STATE.md, changelog and comments updated; redundant tests
trimmed.
The HUD sends its locale with the summary context; the proxy names that
language in the instructions and the input, English when missing or
unsupported. The model returns {summary, feedState}; the HUD shows the
summary only when feedState matches its own state for a view that is not
nominal, and appends its own translated state tag.
partial joins the shared feed-state table between fallback and nominal,
so one partial layer no longer reads off and partial plus nominal no
longer reads nominal.
Japanese in the language picker: the full catalog, counters, unit words joined
to their numbers (5分前, 1時間5分), lower-case HUD units as in Chinese, and the
CJK label rules plus strict line breaking for Japanese. Based on translations
by @abiyeenzo (#913).
Co-authored-by: Enzo Abyie <[email protected]>
German in the language picker: the full catalog, neutral infinitive wording,
one/other plurals and the de style guide. Directions keep German nouns
capitalized, and layer names break at word hyphens but not inside codes.
Based on translations by @abiyeenzo (#908).
Co-authored-by: Enzo Abyie <[email protected]>
French in the language picker: the full catalog, French typography, one/many/
other plurals and the fr style guide. Voice plan steps join through a catalog
string so French spaces its colon, and with nine languages the picker shows
its search field. Based on the translation by @abiyeenzo (#907), checked
against @meirm's French catalog (#312).
Co-authored-by: Enzo Abyie <[email protected]>
Co-authored-by: Meir Michanie <[email protected]>
uk.json: keep both sides' new keys. The new it and pt catalogs get the fire
filter strings; "≥ {value} MW" (it, pt) and "Nominal+" (pt) stay as in
English, like es.
Heat was a detection count, so raising the minimum-FRP slider removed most
detections and the hot spots faded to blue even where strong fires still
burned. Each detection now adds its FRP in megawatts (0.5 MW floor), so weak
fires carry little heat and filtering them out keeps strong-fire areas hot.
On the live payload, hot spots that stay yellow or hotter at >=100 MW rise
from 7% to 17% and their median heat from 0.10 to 0.38.
Brazilian Portuguese in the language picker: 2732 strings, Portuguese plural
forms and the pt style guide. The CCTV summary line names its calibration,
source and view in the page language, markup can give text a context with
data-i18n-context, and the cyclone source errors are translated. Based on
translations by @abiyeenzo (#909).
Co-authored-by: Enzo Abyie <[email protected]>
Italian in the language picker: 2730 strings, Italian plural forms, labels
shortened to fit the panels, and the it style guide. Based on the translation
by @abiyeenzo (#912).
Co-authored-by: Enzo Abyie <[email protected]>
German Khokhlov