Keep four-sample geometry and depth while using single-sample postprocessing buffers. Preserve forward and overlay coverage, skip empty transfers, and add matched pixel and regression checks.
Batch consecutive disjoint installation outlines to avoid three stencil draws
per feature. Preserve overlapping volumes, material order and local precision.
Delete the public effect handle correctly during teardown.
Keep imagery below ground classifications and glyphs in every activation order,
including minimal mode without tone mapping. Exclude fully offscreen ground
volumes from the screen candidate grid.
Full unit/allocation, boundaries, formatting and keyless builds pass. Browser
checks cover contour coverage, bounded colour differences, a missing-contour
negative control, real pass order, sprite alpha and FIRMS glows. Whole-wave
performance acceptance remains separate from this implementation checkpoint.
Clear the display-resource signature used by static sprite reuse when the WebGL context is lost or restored. Cover both events and verify that caching resumes after the redraw.
Port camera markers, coverage, media planes, volumes and calibration to the
Navara renderer. Preserve shared source/media policy and add exact plane
picking, owned resource cleanup and idle hold handling.
Validate paired keyless dev/preview media and physical drags, native legacy,
engine contracts and full CPU/build gates. Record remaining shell, viewshed
browser and W2 performance gates; defer Dioxus and plan the transit port.
Reuse retained tile triangles for CCTV's upcoming activation probe, with finite range, exclusions and honest misses. Correct nearest-hit ordering when the ray starts inside a mesh bound.
Fifteen focused surface tests and all full keyless unit/allocation, boundary, format and build gates pass. CCTV renderer integration and existing acceptance gaps remain documented.
Match native ground footprints, encoded alpha and source ordering on real terrain and photoreal tiles while excluding forward meshes. Cache static screen data and reduce large-disc edge work without changing GPU pixels.
Add deterministic lifecycle, source/share, foreground/horizon and three-context performance proof; restore credits for available Navara feeds and retain failure history in migration evidence.
Render satellite fleets and selected targets through mover banks and orbit lines.
Enable the celestial HUD on Navara and keep native follow explicitly in ENU.
Add deterministic interaction, share, orbit, motion and cleanup acceptance QA.
Record complete targeted acceptance and remaining shared cross-shell tile failures.
Minimal inherited the sun light default of castShadow=true although the photoreal tiles carry baked lighting. Use the same no-shadow sun configuration as cinematic. The focused 3-run city gate improved from C4.6/N5.1 ms (1.1087) to C4.6/N4.9 ms (1.0652) with 12000 vessels. Full Node 24 tests and allocation gates, boundaries, format, build and panel build pass.
Fix delayed and growing mover bank commits, reject stale hidden model picks, and route world-focus through the engine. Strengthen vessel parity and performance checks without changing thresholds.
T11 review follow-up to ed9ea7e6. CPU hits rank as the engine draws them:
sets without a depth test over depth-tested ones, then by set order
(the tracked bank over the fleet, flights over local ADS-B), then depth —
so the CPU's top hit that checks a GPU answer is the instance drawn on
top. A model hit is no longer taken as answered: the model mover keeps
each instance's drawn diameter (after minimumPixelSize/maximumScale) and
the CPU checks the hit against its projected position within that radius
(+3 px); the CPU pick sees models at that radius too.
The instance colour (sRGB, as Cesium takes colours: the military amber
#FFB800) was mixed into the linear material colour as is, so a tinted
model drew paler and brighter than Cesium's: the military footprint gate
read 124 % (more pixels cleared the contrast threshold over water) and
the median luminance +9 %. White (flights) is the same either way, which
is why the flights gate passed. The tint is now sRGBTransferEOTF-decoded,
as the sprite sets already decode theirs.
qa-flights-engines' model gate adds the highlights: the p90 luminance of
the model pixels within ±10 % of Cesium, with the near-white pixel count
reported (10b review low 2).
Models on, near pose (computed scale ratio ~1.0000 on both):
- military (from 100 km): footprint 9,902 vs 8,610 px (115 %), median
luminance 158.9 vs 155.3, p90 190.9 vs 184.5 — 18/18;
- flights: footprint 104 %, median 189.8 vs 196.7, p90 223.1 vs 247.8
(near-white px 83 vs 510: Cesium's white models saturate where Navara's
tone mapping rolls them off) — 20/20.
Models off: flights 16/16, military 15/15. Side-by-sides regenerated.
Navara 0.2.1 can answer a click with the previous gesture's pick. The
check let that answer through when its instance was drawn within twice its
pick radius of the click — the followed aircraft at the centre beside the
clicked one (the 20/24 flights checklist run: a click selected the
previously followed aircraft). A GPU answer for a sprite or point set now
has to be drawn under the click (no slack) and be the instance the CPU
pick finds on top there (lines left out: they are not in the GPU pick);
otherwise the CPU pick answers. Models, which the CPU pick does not see,
are still taken as answered. The CPU drill takes `includeLines`.
Test: 20 gestures alternating a click on A and a click on its neighbour B
answered with A's stale pick — B every time. In the app, 40 scripted
click/follow/release gestures on Navara selected the clicked aircraft
every time with and without the fix (the stale answer is rare).
A sprite/point set uploads its instances by update range. When the camera
went far enough to move the set's origin, update() rewrote every instance
(relative to the new origin) — and a commit later in the same frame
cleared the ranges and kept only its own, so the GPU never got the
rewrite: the instances not committed again drew offset by the origin move
(a military glyph at the orbit pose drew 45 px from its point after the
global → city → orbit sequence; the CPU picks, the probe and Cesium had it
in place). Ranges now add up until three uploads them (it clears them
after the upload).
Found by the military side-by-side (11b); flights' sprites had the same
exposure since 10b.
The near-pose finding (Navara models faint, Cesium ~23 bright aircraft)
came from the side-by-side's frozen page clock: the fleet tick ran once per
pose, so Cesium never handed its aircraft to their models and drew 3×
glyphs, against real models on Navara. With the hand-off driven (the move
ends every 10 frames while a pose loads), both engines draw models, and
the comparison showed three real differences, all fixed:
- Orientation: Navara's glTF loader is three's GLTFLoader, which keeps
the scene Y-up; the instance turn is now Cesium's whole axis correction
(Y_UP_TO_Z_UP × Z_UP_TO_X_UP), not Z_UP_TO_X_UP alone. Every fleet model
drew 90° off its heading.
- Size: minimumPixelSize is taken at Cesium's distance (Camera
distanceToBoundingSphere: view depth less the radius at the model's own
scale), in the mover and in ModelHandle.computedScale; the straight-line
distance drew off-centre models up to 16 % larger.
- Colour: the instance colour tints by the set's blend amount (Cesium's
colorBlendAmount) and its alpha is the model's opacity (Cesium: a
model colour's alpha makes it translucent), not a blend weight; a model
is translucent only while faded. An ambient share of the tinted surface
(MODEL_AMBIENT 0.5, Cesium's image-based lighting) keeps a tinted
aircraft near-white at a low sun (median luminance 139 → 190; Cesium 197).
qa-flights-engines (now --layer=flights|military) gates the near pose
with --models=on: computed scale within ±5 % per model (ratio 1.000007),
footprint against the same view with the layer off within ±20 % (1553 vs
1496 px, 104 %), median luminance within ±10 % (189.8 vs 196.7). The probe's
drawn() reports a model's computedScale. Flights models on: 19/19.
FollowHandle.target() returns the followed point, checked by the contract
on both engines. The contract's model check now waits for ready() before
looking for the model's pixels: the magenta point and sprites could fill
the region before the glTF loaded (qa-globe-contract cesium 73/73,
navara 79/79).
Linear, without mipmaps: Cesium's billboard atlas has none, and a mipmapped
atlas greyed small glyphs Cesium draws at full colour (a 5 px aircraft at
15,000 km: 239 against Cesium's 255; now 255 on both). Task 10b colour
parity.
Task 10b foundation.
- MoverSet.ready(): whether what the set draws has loaded (Navara: the
model set's glTF, a URL atlas; Cesium: every instance's Model, the cropped
atlas; parts without it draw at once). The aircraft renderer keeps the 2D
glyph until a model set is ready, as the Cesium path waits on Model.ready.
Contract suite: a drawn model set is ready() (both engines); unit test of
the neutral store.
- Navara sprites and points interpolate scaleByDistance as Cesium's
NearFarScalar (over the squared camera distance, eased by ^0.2) instead
of linearly in distance, so aircraft glyphs have Cesium's size at every
range; the CPU pick radius follows (distanceScale, tested equal to
focusDeemphasis' port of czm_nearFarScalar). The contract text says so.
Phase-4 QA: after a map-stack switch at the whole-globe view Navara
sometimes showed the new map only seconds later. Navara 0.2.1 reports idle
and stops drawing before a raster tile it fetched arrives, so the last
frame stays stale until something else asks for one. Every engine.maps
change (base imagery, terrain, photoreal tiles, their release) now asks for
a frame and keeps asking every 250 ms for 6 s, then the engine idles again.
scripts/qa-map-switch.mjs switches the stack (photoreal / Esri / OSM) at
18 000 km on each engine, waits for the tiles, hashes the picture, forces
frames and hashes again: a change means a stale picture. Before: Navara 3/8
switches drew the new picture by themselves; after: Navara 8/8, Cesium 8/8.
Review of Tasks 6-7: a 240 px drag right turned Navara's follow by -67.5°
where Cesium turns +119.5°, and a right drag did nothing (Cesium dollies,
1.95 Mm → 0.91 Mm). followGestures.js now follows Cesium's
ScreenSpaceCameraController for a tracked camera:
- left drag (rotate3D): heading += 1.77 × dx / width × 2π (clockwise for a
drag right; 1.77 is Cesium's maximum rotate rate, which a tracked camera
beyond a few metres always reaches) and the camera rises by
1.77 × -dy / height × π; each move counts at most a tenth of the canvas
(maximumMovementRatio);
- right drag (zoom3D): follow.js dollyOffset moves 5 × dy / height of the
distance above the floor (closing) or of the range (opening), never
inside it; no context menu while following.
The contract follow check turns the camera with a 240 px drag right and
asserts the heading turn against Cesium's rate (±15 %; the harness holds the
pointer before releasing it so Cesium adds no inertia), and dollies with a
right drag down (range below 80 %, never under 149 m). Cesium 73/73, Navara
79/79.
The shell creates detection and the tracked readout on the engine, so the
Navara mode has them too (before: not created without a viewer). Detection
reads the camera height from the engine. The tracked readout follows
engine.camera.following(): a follow's plain-data label
(FollowOptions.label: { id, model, position }, the readout card's model
and anchor) or, on Cesium, the followed Entity and its gevLabelModel as
before (Cesium layers keep publishing the same). refreshTrackedReadout
takes a follow handle too. A bare viewer without a registered engine
(tests) is still read for its trackedEntity.
In the app on Navara: "[Detection] Initialized", "[TrackedReadout]
Initialized", and a follow with a label paints its card.
The world overlay host already projects through engine.projector and tests
the horizon with projector.horizonVisible; the last Cesium value on its
path, the snapshot of an entry's cullPosition, is now a geo Cartesian3.
Navara's horizonVisible is Cesium's EllipsoidalOccluder.isPointVisible:
tested against Cesium's own on 8 000 eye/point pairs (600 m to 20 000 km,
points under, on and over the ellipsoid; no disagreement) and at the limb
(a label 40 m under the ellipsoid just past the horizon is hidden, with its
cullPosition lifted 100 m it is shown). It ran per label per frame with
three new vectors each call; it now uses scratch vectors. Micro-measure
(500 labels per frame, no scavenge during the run): 7 941 B per frame
before, 51 after.
The W1 sources that projected with Cesium's SceneTransforms themselves -
the icon orientation's exact projection and the launch replay overlay - go
through the engine's projector when their scene or viewer has a registered
engine. Satellites, vessel labels and aircraft recession project nothing
themselves (the host does).
scripts/qa-overlay-engines.mjs publishes the same labels on both engines at
a city, an orbital and a limb pose and compares the painted rectangles:
worst 0 / 2 / 0 px; on both, the far-side label is not painted, the lifted
limb label is and the buried one is not.
Review of Task 4: the CPU picks (Navara hover, drill, the stale-answer
check) took a sprite or point at its full sizePx however far it was, so a
set shrunk with scaleByDistance was picked well off its drawn edge. The
radius now scales by the set's scaleByDistance at the instance's camera
distance, linear between near and far as the Navara shader draws it.
Review of Task 4:
- a Cesium pick whose top result is filtered (a pickable: false set, the
follow camera's proxy entity) took nothing; it now drills and answers the
first result under it with an id (an object), else the first kept one
(map tiles);
- billboards and points drawn without a depth test reach Cesium's GPU pick
through the globe: mover instances behind the horizon (projector
horizonVisible on the instance position) are left out of picks and
drills, as Navara's CPU drill already did;
- the follow proxy entity (1 px point, cesium/follow.js) carries
gevFollowProxy and is never a pick (review item 6).
Review of Task 5: while following on Navara a left drag panned the globe
about the ground point under the cursor (range 2,154 m → 54,136 m) and the
wheel zoomed toward that point (462 m after 40 notches, Cesium stops at
the 150 m floor). While a follow runs, Navara's spin, zoom and tilt are
now suspended (input.suspendCameraInputs, which gives the app's own
setCameraInputs choice back after the release) and followGestures.js turns
a drag into heading and pitch of the ENU offset (a canvas width is one
turn, its height half a turn; elevation kept within 2°-89°) and a wheel
notch into 20 % of its range, never inside the minimum range, as Cesium's
EntityView orbits the target. The offset is no longer read back from a
camera the follow alone moves.
Also (review item 7): a follow requests a render only when the target, the
offset or a gesture changed, so a still target lets Navara idle.
The contract follow check drags (range within ±5 %, target within 3 px)
and turns the wheel 40 notches (never below 149 m, ends within 5 % of
150 m) on both engines with real input (qa-globe-contract's pointer gains
drag and wheel); after stop it waits for Cesium's gesture inertia to end
before checking that the camera is free. Contract: Cesium 73/73, Navara
79/79.
Review of Tasks 4-5: the overlay pass encodes linear to sRGB, and the
instance and line colours went in as if linear: grey 0.5 drew 187 (Cesium
128) and [1, 0.6, 0.2] drew 255/203/124 (Cesium 255/153/52). The sprite,
point and line shaders now decode them (sRGBTransferEOTF), as the sRGB
atlas texture already was.
A contract check draws mid-tone sprites, points and lines (grey 0.5 and
[1, 0.6, 0.2]) and reads their centres: within 3 per channel on every
engine. Navara 79/79.
engine.surface.sampleHeightDetailed(place, { exclude, lod }) answers
{ height, source: 'tiles' | 'terrain', tileSizeM? } (src/globe/surfaceDetail.js).
On Navara tileSizeM is the extent of the tile that answered and
lod 'finest' takes the finest loaded tile, drawn or not; Cesium answers
tiles without a size, else the globe's terrain. acceptedSurfaceHeight()
turns an answer into a height or a miss: tiles larger than
MAX_FLOOR_TILE_SIZE_M (4 km) are too coarse, terrain is not the mesh.
The mesh floor sampler and the ground snap are engine-only now (a viewer
or scene reaches its engine through engineFromViewer, which also registers
the scene) and treat a coarse answer as a miss, as they treat undefined.
Measured on Navara at Austin: while the tiles stream, a 27.9 km tile reads
143.2 m where the ground is 122.5 m (a 1.8 km tile).
`exclude` takes the engine's own mover sets (MoverSet.id): Cesium hides a
set's primitive for the synchronous sample (Cesium matches excluded objects
per billboard, not per collection); Navara's ray only meets tiles.
"sampleHeight undefined in the app on Navara" did not reproduce: the
startup Austin sample is 121.9 m (tile 889 m) on Navara and 122.4 m on
Cesium; one ModelMesh class is patched.
S2 QA (scripts/qa-surface-heights.mjs, 20 places at JFK, SFO, AUS, roofs
and streets): Navara within 2 m of Cesium at 19/20 from 800 m and 18/20
from 15 km with lod 'finest' (gates 18 and 15); offsets per region
-0.21 / +0.4 / +0.1 m; the Empire State roof differs by 39 m (spire
geometry). Navara tile sizes at 800 m: 112-314 m.
camera.follow({ position, viewFrom, minRange = 150, native? }) keeps a
moving target centred at an east-north-up offset the user may turn and
zoom; camera.following() and camera.followChanged report it; stop() frees
the camera and a camera command (setView, flyTo, flyToSphere, lookAt,
cancelFlight) ends it.
Probe ruling: Navara 0.2.1's cameraFollow holds the target but ignores the
user's drag and wheel, the aircraft sits ~13 px off centre with 7.6 px of
spread, and calling it without an offset each frame gives a NaN camera.
So Navara (and the fake engine) follow with src/globe/follow.js: each
frame, before Navara draws, the offset is re-read from where the user put
the camera relative to the previous target, kept outside the minimum
range and carried to the new target.
Cesium follows with viewer.trackedEntity: the layer's own Entity
(`native`) or a proxy Entity at position() with a 1 px transparent point,
since EntityView moves the camera only for an entity with a bounding
sphere. applyTrackedCameraFrame and the zoom limits moved to
src/globe/cesium/follow.js; src/data/trackedCamera.js keeps the pure parts
(the range clamp now lives in src/globe/follow.js). following() also
reports an entity the app tracked itself. Camera sensitivity claims made
on a viewer's Cesium camera now share the engine camera's ledger, and the
orientation controls and the world jump read the follow from the engine.
Contract (both engines): a target flying at 250 m/s stays within 3 px of
the centre for 2 s, zoomIn through it stops at 150 m, stop() and setView
end the follow. Cesium 71/71, Navara 77/77; shell 23/25/4.
Review of Tasks 2–3:
- the mover store's instance checks no longer build arrays or closures
per set() (orientation and colour checked field by field), and dirty()
returns one reused range object;
- Cesium models get their matrix and colour through scratch objects
(Model.modelMatrix and Model.color copy what they are given) instead
of a new Matrix4 and Color per placement;
- a destroyed MoverSet releases its store (count 0, empty instance
arrays, no pick ids) so a set object a layer still holds keeps no
memory, and the Cesium parts drop their billboard, point and model
lists;
- (with the overlay commit) the Navara line viewport is updated in
place.
Node micro-measure, 10 000 instances set + packed per frame: 0.27-0.29
ms, about 1.7 KB per frame (0.17 B per instance), as before. Contract:
Cesium 70/70, Navara 76/76.
Review of Tasks 2 and 4:
- A contract check draws, on both engines: an atlas cell switch, a
rotation of π (a half-white cell flips from above to below the
centre), scale 2 (about four times the area), scaleByDistance past
`far` (about a quarter), setCount shrinking the set, depth modes for
sprites 30 km under the ground ('always' and a disableWithinM past the
camera draw), a point with its dark outline, and a glTF model
(airplane.glb, minimumPixelSize 80, colorBlend replace) in magenta.
'normal' and a 10 m disableWithinM hide the buried sprite on Navara;
Cesium's billboards are not occluded by the app's photoreal tiles
(globe hidden; measured), so those two are asserted on Navara only.
- pickable: false: Cesium's GPU pick sees every primitive, so the
picker drops picks (and drill entries) of a set created with
pickable: false; Navara only registers pickable sets. The contract's
click check now clicks a pickable: false sprite on both engines.
- Navara resolves a GPU answer's batch id to the pick id when Navara
answers (the pick's own frame), not when the engine's click consumes
it: a set that reuses the instance in between still yields the
instance that was under the pointer.
Contract: Cesium 70/70, Navara 76/76.
Review of Task 3 (DPI-independent dashes; the shader change landed with
the overlay commit): qa-globe-contract --dpr=2 runs the suite on a
high-DPI page, and a new check measures a 16/16 px dashed line along its
middle row in CSS pixels: the dashes must be regular, and on Navara
16/16 within 2 px. The harness prints what each engine draws.
Measured: Navara 16/16 (period 32) at DPR 1 and 2. Cesium draws
16 on / 16 off (PolylineDash dashLength 32, pattern 0xff00, uniforms
checked on the primitive) as 14 on / 2 off with a 16 px period at both
ratios — a Cesium-side difference to look at before a layer relies on
pixel dashes (none does yet). At DPR 2 the Navara harness's
stars-at-midnight check counts 4 star pixels (its detector is tuned for
DPR 1); every suite check passes at both ratios.
Review of Task 4: the CPU line pick (shared with the sprite drill, hover
and the synchronous pick) took half the drawn width + 3 px, so a 1-2 px
trail needed a 3.5-4 px aim. Now within 8 px of the projected segments,
or half the width + 3 px for wider lines. Tested at 7.5 and 8.5 px from
a 1 px line; the contract suite's line click passes on both engines.
Review of Tasks 2–3: on Navara the overlays went through the look's
tone mapping and grading (magenta came out as 247/40/228, green as
147/228/89 in the cinematic look), where Cesium draws billboards and
polylines in the colours asked for.
src/globe/navara/overlay.js: a public Navara EffectDesc (gevOverlay)
inserted after toneMapping and before the gev sharpening, SMAA/FXAA and
final, whose pass renders one three scene over the tone-mapped input
buffer (in Navara's own Pass wrapper: Navara inserts a bare pass only
when it is an instance of its own postprocessing copy); it costs nothing
while the scene is empty. The chain has no depth buffer, so each overlay
material tests the scene depth itself (overlayDepth.js): its log depth
against Navara's MRT depth (RGBA packed), modes always / normal /
depth-fail, disableWithinM as a per-vertex escape. Sprites and points
(moverMesh.js) and line ribbons (lineMesh.js) live in that scene with
their per-frame work as overlay hooks; models stay Navara meshes, lit
and graded like the tiles. The sharpening now sits after the overlay
(Cesium's sharpen stage sharpens billboards too).
Also (review): Navara pixel dashes are measured in CSS pixels
(gl_FragCoord over uGevPixelRatio), and the line viewport uniform is
updated in place.
Measured: a magenta sprite draws 255/0/255. The contract suite's colour
checks are now within 40 of the asked colour on every channel on both
engines (were: dominant channels); Cesium 68/68, Navara 74/74 (its
sharpen-order check accepts the overlay between tone mapping and the
sharpening). Shell: cesium 23/23, cross 4/4; navara 23/25 — the two map
stack pictures from space fail on stddev 7.2/7.6 against 8, as on the
Task 3 commit run at the same hour (the cinematic sun follows the clock
into night over the reset view), not from this change.
Navara 0.2.1's transparent pass refreshes its copy of every scene light
each frame with copy(light, true). For the sun's cascaded shadow lights
(three-csm CascadedDirectionalLights) that copy clones every cascade, so
each frame brought new DirectionalLights with new ids and three's
WebGLLights uniforms cache (keyed by id) grew by about 1.2 KB a frame in
the looks whose sun casts shadows (minimal, photoreal).
src/globe/navara/cascadedLightCopy.js patches the copy to update the
cascades it has in place (pose, colour, intensity, target, shadow
settings and camera) when the counts match, and keeps the rebuild
otherwise; internals.js finds the class among the view's scene lights
(private view._scenes.light) and applies it when the globe is built.
Tests on three-csm's own class: the original copy changes the light ids,
the patched one keeps the lights and takes the values.
Post-GC JS heap (CDP HeapProfiler.collectGarbage every 10 s, the app on
?engine=navara rendering continuously at 60 fps): minimal 4.33 → 0.20
MB/min over 70 s and 0.06 MB/min over 180 s; photoreal 4.34 → 0.19
MB/min; cinematic (no shadow) 0.34. Added to the upstream asks.