The shared launcher always asked for ANGLE's Metal backend, which exists
only on macOS; on CI's Linux runners WebGL failed and the Street Level gate
never saw the app start. The default is now Metal on macOS and SwiftShader
elsewhere; QA_ANGLE still overrides it.
The tileset's transform and probe points come from trigonometry whose last
bits differ between arm64 and x64, so a fixture written on one failed the
byte comparison on the other. The numbers now match to 1e-12 of their size;
the meshes and the served files are still compared byte for byte.
The tile budget left 768 MiB for tiles over the screen buffers, while an
oblique view needs about 0.9 GiB at 960x640 and 1.2 GiB at 1440x900; over
budget Navara coarsens the view and stops loading, so the start view stayed
at imagery level 15 where Cesium shows 17. The room for tiles is now
Navara's device default (a quarter of device memory, at most 2 GiB).
Terrain tiles also split past the keyless mesh's level 14, so imagery can
refine beyond level 15, and OSM stops at its level 19.
Esri answers a tile it has no imagery for with a grey "Map data not yet
available" image (over open ocean from level 14). Cesium discards it and
keeps the parent level; on Navara that tile now fails the same way.
Both probes called waitForEngine without the engine kind, which can never
match and timed out after 120 s; waitForEngine now rejects a missing kind.
The trail probe passed no result vectors to the geo helpers, and drew its
trail below the ground, where only the dimmed occluded pass showed it; it
now draws halfway down to the ground and both engines count the same
pixels.
Translate the 148 strings main added since (mostly the News layer) and
prune the 5 it no longer uses; keep both Hebrew and Swedish in the
language lists.
Brings in 190 commits from main. Integration on the engine contract:
- News and Shipwrecks draw, pick and keep their selection through the
contract and shared map markers, so they follow engine switches.
- Aircraft tracking uses main's segmented trails with contract frame
numbers.
- Decoded weather tiles move into the Cesium adapter and are injected
through the renderer registry; Navara hides those sources on globe
surfaces (docs/KNOWN-ISSUES.md).
- Street-vehicle GPU depth occlusion stays Cesium-only; Navara keeps
the geometric visibility checks (docs/KNOWN-ISSUES.md).
- Transit includes the heading-neutral atlas icons.
- The cloud brightness key is a filter, not a brightness scale.
- Counted lines without a plural form read "label: count", so 1 agrees.
- Fire confidence is «tillförlitlighet», not «säkerhet».
- Agreement with «flöde» and «luftfartyg»; course on the tracking line is
«FÄRDRIKTNING», which fits both aircraft and ships.
- Wildlife is «Artobservationer», since the layer also shows plants and fungi.
- Bridge readouts are nouns (BLICK, ÖGONHÖJD), cockpit key hints share one
verb form, tool names in pointer messages are quoted, and «jfr» replaces
«jmf.»; the guide's example follows.
- Smaller wording and consistency fixes across the catalog.
GDELT exports are read in parallel, so two reports with the same time
arrived in whichever order their exports finished, and the feed test that
pages through them failed now and then. Same-time reports now order by
most mentions, then by id.
A weather feed may also be a frame source. With an optional
getFrameImage({ product, time, size, bbox, signal }) the layers take
whole-extent images, detail windows and the global mosaic from it; with an
optional getFrameTile({ product, time, z, x, y, size, signal }) regional
tiles and their prefetch come from it, null being a transparent tile. Pixels
may be an ImageBitmap, ImageData, canvas or VideoFrame of exactly the
requested size; the layer draws them once and closes them. Cancellation
reaches the source, and a feed without these methods reads the image and
tile routes as before.
AIS Vessels are keyed by the id their live source gives, with the MMSI as an
attribute: an MMSI can be shared, spoofed or reused, so two vessels with one
MMSI are now two vessels. Looking a vessel up or selecting it by MMSI still
finds it.
A share link made while a vessel is selected opens that vessel. A UUID vessel
id travels as 16 bytes in a new thing field (13). A shared thing that is not
reporting now says so in a sentence per kind, with its id (ICAO, NORAD, MMSI
or vessel UUID) on a second line.
A live source may now declare a timeline: it can answer a snapshot at a
given time (source.timeline = { coverage(), stepMs, subscribe? }, read by
readSnapshotTimeline). While such a layer is on it follows the world clock
and registers the coverage. At the live edge it polls exactly as before;
away from it the layer asks for getSnapshot({ ...query, atMs }) at the step
at or before the clock time and shows that snapshot instead of polling.
src/time/snapshotTimeline.js owns the following: a seek, time going
backwards or leaving live replaces every contact with the snapshot at the
new time (play waits on it, for at most 8 s); playing on keeps one
snapshot ahead and gates play on it, applying those snapshots as polls so
contacts move smoothly; outside coverage the layer is empty; back at live
it is cleared and polls at once. One load at a time, and a load the
source drops for a newer one is not a failure.
While following, aircraft are drawn at the clock time with no render
delay, records and absences are judged at the clock, track requests carry
atMs, and vessels move between fixes along the great circle. The world
clock counts jumps (seeks and returns to live) in getState().
A source with no timeline never follows the clock, and every layer stays
live, unchanged.
The radio and push-to-talk helpers shared by both voice providers now pass
English catalog text, and the OpenAI controller words its detail when shown,
like the Gemini session. A token failure without a server message carries
its status as sourceParams, so it is worded again with the number. The docs
say sourceDetail is the message as given, catalog text when it is the app's
own.
The Gemini session translated each status and error message once, when it
happened, so after a language change the voice control kept showing it in
the old language. The state event now also carries the English catalog text
as sourceDetail, and the control words it when shown. The local server's
known token errors stay catalog text until then instead of being translated
on arrival.
The Gemini token route now answers a missing key with code API_KEY_MISSING,
the session carries it to the error state as missingKey, and the voice
control hides "Check microphone permission" for that error, as it already
does for OpenAI. The OpenAI controller sends the same flag, so the shared
control and the controller agree whichever updates last.
The voice error was matched against the translated key message to decide
whether to show the microphone hint, so switching language brought the hint
back and left the message in the old language. The error now carries a
missing-key flag into the status, which words the message at render time.
A controller test covers the language switch, and CURRENT-STATE.md documents
the keyless voice response.
The decision no longer reports not-entity-context or channel-closed, which
its only caller rules out first, nor pointer-scale-not-useful, which the
crop path already checks and logs. A closed channel returns before any
record, as on main. The crop decision is pointer-crop-eligible, since the
crop checks after it decide whether the pointer is on ground, and a sender
called on a closed channel logs channel-lost.
In a dense downtown view from about 400 m up, the occlusion margin of
max(3 m, 2% of the range) came to 14 to 20 m, so a wall, awning, tree or
lower facade standing 10 to 20 m in front of a street was read as the
road itself: the vehicle's dot was hidden by the scene, but its bracket and
callout were still drawn. Checked against scene.pickPosition in one such
view, 18 of 470 vehicles in frame fell in that band.
The margin now depends on how steeply the camera looks down on the
contact. A point on the line of sight s metres in front of a vehicle
stands s * sin(elevation) above it, so a surface there hides the vehicle
once that rise passes 2 m. The margin stays at least 3 m and at most 2%
of the range, so grazing and street-level views are unchanged. In the
same views no vehicle with the drawn surface within 3 m in front of it
(its own road) is newly hidden.
From the review of the first change:
- A contact back in view after longer than a retest interval is held back
and retested like a new one, instead of being drawn from an answer
given before it left the view. It gets a fresh record, so a test still
in flight from before cannot end the hold-back either.
- The hold-back for a contact's first answer stretches with the retest
pace when a crowded view slows it.
- A batch whose control point failed no longer counts as a first answer.
- An unknown log-depth flag reads as "no depth" rather than as standard
depth, which would have put every surface in front of every contact.
- A shader that fails to build leaves no shader behind and is not built
again on the same context.
The detection overlay only hid a contact behind the Earth's curve, so a
street vehicle behind or under a building, a 3D tile or a hill still had
its bracket and "VEH-nnnn" callout painted over that geometry, while its
own dot was depth-tested and correctly hidden.
A contact a layer marks `depthTested: true` (only Street Traffic does) is
now tested against the depth the scene drew and left out of the overlay
while a drawn surface stands in front of it by more than max(3 m, 2% of
its range). Other layers are untouched: their markers are drawn over the
scene on purpose.
The test runs on the GPU and never waits for it. Each tested contact is
one point drawn inside an occlusion query; the vertex shader looks up the
frame's packed pick depth at the contact's pixel, decodes it and drops the
point when the surface is nearer than the contact's threshold. Query
results are collected on later frames. Reading depth back was measured
and rejected: `scene.pickPosition` per contact costs 0.2 to 1.4 ms each,
and Chromium answers `getBufferSubData` with a round trip to its GPU
process even behind a signalled sync object (1.5 ms on a quiet machine,
15 to 25 ms on a busy one).
- Contacts are retested every 100 to 200 ms, spread across frames, at
most 256 a frame; with more in view the pace stretches so none starves.
- Hiding is immediate. A hidden contact returns 600 ms after its last
"hidden" answer, so one on a silhouette edge does not blink, and a
contact new to the view is held back until its first answer (400 ms at
most), so orbiting does not flash brackets behind buildings.
- A callout whose contact goes behind something is dropped at once, not
carried to the next solve and not faded out on the facade.
- These times run on painted frames, so a hidden tab does not expire them.
- A control point in every batch must draw; if it does not, the batch is
ignored and after three the probe stands down for ten seconds. With no
pick depth (2D, Columbus view, depth picking off) nothing is hidden.
A WebGL 1 context reads depth back synchronously about 8 times a second.
The renderer internals this depends on are read in one function,
`readSceneDepthSource`, with the engine version they were checked against.
Roads the view cannot sample may keep their pending surface heights, as
before the refactoring; the oblique wait now asks for current coverage,
drawn traffic, finished loading and no scheduled refinement, and prints
every loading field when it times out.
Each refinement pass renewed the ten-second deadline, so changing
surface readiness could keep the layer refining, and loading, without
end. The passes now share one deadline, and the scheduled surface work is
reported as surfaceWaiting.
Field entries in the ledger allowlist now apply to every difference row,
in the ledger and in the probe alike, and may carry maxDelta. The traffic
fill gets one with a bound of three: its simulated vehicles enter and
leave every frame, so two pages snapshotted moments apart differ by a
vehicle or two.
German Khokhlov