@nika-js/onlymap 0.7.5 → 0.8.0

package/llms.txt

@@ -34,7 +34,7 @@ Programmatic/native bridge rule: `MapController.setLayers()` accepts normal func
 ## Element vocabulary
 
 - `<om-map center="[lng, lat]" zoom="11" pitch="55" bearing="20" basemap="positron">` — the root. Give it a height (`om-map { display:block; height:100vh }` with `html,body{height:100%}`, or a sized container) — a custom element is display:inline by default and collapses to zero size; with none set, the library falls back to display:block + a 400px floor so a bare map still shows (any height you DO set wins over the floor, including one below 400px), and a collapsed map logs a console warning. A map hidden on purpose (`hidden`, or inside a `display:none` panel) stays hidden and does not warn. `basemap` accepts a free preset (`liberty`, `bright`, `positron`, `dark-matter`, `voyager`, `osm` — no keys; or `maptiler-streets`/`maptiler-dataviz`/`maptiler-satellite` with `basemap-key="…"` or `OmMap.configureBasemap({ maptilerKey })`), `maplibre` (bare demo style), a style URL (e.g. a MapTiler-customized style; any scheme fetch supports, including desktop asset protocols like Tauri's, query strings fine), or `none` (standalone canvas). The attribute is LIVE — writing it switches the basemap in place (camera + layers survive); the `set-basemap {basemap}` action and `<om-widget type="basemap-switcher" options="positron dark-matter osm">` do the same. Register more via `OmMap.registerBasemap(name, { style })`. Required provider attribution renders automatically (`attribution="false"` opts out). `pitch`/`bearing` tilt the initial camera (use for 3D content). Terrain: `terrain="terrarium"` (keyless AWS DEM; also `mapterhorn` — keyless, CARTO Positron drape by default — `maptiler-terrain` with a key, a raw `{z}/{x}/{y}` DEM URL + `terrain-decoder="terrarium|mapbox-rgb|<JSON>"`, or `off`) raises a 3D surface — geographic layers DRAPE onto it automatically (per-layer `terrain="drape|offset|off"` overrides; 3D-model layers sit ON it), `terrain-exaggeration` scales relief, `terrain-texture` drapes imagery; terrain REPLACES an active basemap while on (restored when off); `set-terrain` action, `terrain` watch token, `OmMap.registerTerrain(name, {...})` for more sources. a georeferenced BIM model REQUIRES the map to author `terrain` explicitly — any value including an explicit `terrain="off"` (flat-ground siting); a model that resolves real elevation (IfcMapConversion + OrthogonalHeight) on a map with no `terrain` attribute raises an ERROR through the validation channel at load time (the library never writes attributes for you — no auto-terrain). Scene lighting for 3D content: `lighting="daylight|studio|flat|custom"` (absent = deck defaults) with `lighting-ambient`/`lighting-sun`/`lighting-sun-azimuth`/`lighting-sun-elevation`/`lighting-camera` overrides and `lighting-sun-date` (ISO/epoch — solar-position sun computed at the map center, wins over azimuth/elevation); attribute-backed (undoable, live-editable), story-steppable via the `set-lighting {lighting, sunAzimuth, ...}` action (`lighting="default"` removes the attribute set; a bare preset is a clean reset — stale overrides clear), native UI via `<om-widget type="lighting">` (preset radios + tuning sliders), widget watch token `lighting`. Clip box (issue #34): `clip-box-min="[lng,lat,elev]"` + `clip-box-max="[lng,lat,elev]"` cut a real axis-aligned 3D box through the scene — every layer clipped by default (per-layer `clip="off"` opts out), `clip-box-invert` shows outside instead of inside, `clip-box-highlight` dims clipped-out geometry instead of discarding it; works on ANY layer including georeferenced Tile3DLayer/BIMLayer content; attribute-backed (undoable) via `set-clip-box {min, max, invert?, highlight?}` (`{clear:true}` removes it), native UI `<om-widget type="clip-box">`; v1 axis-aligned only. XY snapping (issue #34 Part A): `snap="vertex edge midpoint"` + `snap-tolerance="12"` (px, default 12) refines a click/hover to the nearest vertex/edge/edge-midpoint of whichever feature deck ALREADY picked under the cursor — not a spatial index, only that one feature's own geometry is searched, on the CPU, when snapping is on. Applies to every layer by default (`snap="off"` on any `<om-layer>` opts it out, mirroring `clip="off"`) and to a `BIMLayer`'s own edge/crease overlay (its real wall corners/edges, converted from the model's local mesh coordinates to real `[lng,lat]` automatically) — the raw triangle MESH itself is not yet a snap target. Vertex beats midpoint beats edge on range conflicts; hold Space to place a point nearby without snapping. GeoJsonLayer extrudes polygons declaratively: `extruded get-elevation="$height"` (+ `wireframe`). `widgets-hidden` attribute (or the `set-widgets-visible {visible}` action / `<om-widget type="widgets-toggle">` button) hides every widget WITHOUT destroying state — attribution never hides (license); transient (not an undo step) but story-steppable, so a step can clear chrome for a cinematic take. Slots auto-dim while an open `<om-overlay>` popup covers them (position stability over the popup dodging; `widgets-dim="off"` disables). `validate` attribute enables live validation + on-page error panel. Include a `map-id="<random UUID>"` on complete pages (identifies the map artifact for usage telemetry — not a visitor id; the page author deletes it to opt out); `telemetry="off"` disables usage telemetry + library-error reporting for the map (global: `OmMap.configureTelemetry({ disabled: true })`; schema: docs/telemetry.md). Free plan limits (HOSTED http(s) pages only — a dev context (localhost, file://, any non-web scheme) lifts every cap while the attribution badge stays; the exemption is technical convenience, not a license grant — commercial deployment incl. packaged apps still requires a key): 5 layers, 25k rows per layer — layers past a limit don't render and validation says why; a `license-key="om_live_…"` attribute (publishable, origin-restricted) or `OmMap.configureLicense(key)` lifts limits and removes the corner badge. When a user's map exceeds free limits, say so and point at the validation message rather than silently restructuring their data. Events on the element: `om-map-ready` (boot complete; `await mapEl.ready` is the promise twin), `om-validation-error`, `om-view-changed` (camera settled after a move, debounced; `detail = {longitude, latitude, zoom, pitch, bearing, origin}` where `origin` is `"user"` for gesture bursts vs `"programmatic"` for API/story moves — the camera-persistence hook, and the echo-suppression signal when syncing camera state to an app store), `om-map-point` (`detail = {coordinate: [lng,lat]|null, kind}` — every click/hover map coordinate incl. empty-map clicks, for custom capture tools beyond the draw widget), `om-tileset-load` (`detail = {layerId, tileset}` — a Tile3DLayer's live deck Tileset3D, for tools needing the real tileset like region export), `om-tiles-warmed` (`detail = {tilesets, viewports, ms, timedOut}` — a `warm-tiles` pass finished; see the story bullet), `om-route-resolved` (`detail = {layerId, route}` — a Route layer resolved, direct geometry or provider round-trip; `route` carries the normalized geometry/distanceMeters/durationSec/legs/bounds — how a page reads a provider-computed route without re-fetching, re-fired on every re-resolve). `MapController` mirrors these as `onViewChange`/`onMapPoint`/`onTilesetLoad`/`onRouteResolved` options. `await mapEl.snapshot()` returns a canvas-only PNG dataURL of the scene (basemap + layers at device pixels; `{as:"blob"}` for files) — DOM widgets/overlays/attribution are NOT captured, so exports must render provider credits themselves.
-- `<om-layer id="..." type="ScatterplotLayer" data="./points.json">` — any deck.gl layer class by `type` (39 layer types total: 32 bundled deck.gl core/geo/aggregation/mesh layers, plus the native `COGLayer`/`ZarrLayer` raster types, `ImageOverlay` for georeferenced drone JPEGs, `BIMLayer` for BIM source files, and `Route`/`Tracking` for routing/live-tracking — see the dedicated bullet below), plus `PopupLayer` (WebGL badges/labels at scale: `layout="badge|pin-label|card"`, `min-zoom`/`max-zoom`). Every type's full attribute list ships in the package's `onlymapjs.html-data.json` — consult it instead of guessing attribute names or reading the minified dist. TextLayer's default font atlas covers ASCII only: set `character-set` when label text carries other glyphs (`—`, `·`, accents) or deck warns and renders them blank. External layer classes register via `OmMap.registerLayer({type, deckClass, props})` — build them on `@nika-js/onlymap/deck` (the bundled `CompositeLayer`/`TileLayer`/… re-exports), never a separately-installed deck.gl (different class hierarchy, breaks in the renderer); function-valued props ride the subclass's `static defaultProps`; register at module top level BEFORE the manifest mounts (see docs/custom-layers.md). Data: `data` URL (JSON, GeoJSON, CSV/TSV `.csv` — parsed to typed columns — Shapefile `.shp` (+`.dbf` attributes) and KML `.kml` as GeoJSON features, GPX `.gpx` (waypoints/tracks/routes → features tagged `_gpxKind`; a `#waypoints`/`#tracks`/`#routes` URL fragment selects one part), FlatGeobuf `.fgb` (cloud-native binary vector, whole-file decode), GeoParquet `.parquet`/`.geoparquet` (cloud-native columnar vector — all-Point files stay columnar like Arrow, lines/polygons become GeoJSON features; requires the file's `geo` metadata with WKB geometry, and CRS84/EPSG:4326 — a projected CRS is a loud error telling you to reproject, not a silent misplacement; snappy/gzip/zstd row-group compression handled), CityJSON `.city.json` / CityJSONSeq `.city.jsonl` (semantic 3D city models — 3DBAG, PLATEAU — decoded to one of two shapes by the `data` URL, no CityJSON layer type: default → extruded footprints, `type="GeoJsonLayer" extruded get-elevation="$roof_height"`; `?om-surfaces=1` → one row PER FACE at its own real per-vertex height so a pitched LoD2.2 roof actually looks pitched, `type="SolidPolygonLayer" get-polygon="$polygon" full3d` (`extruded` stays at its ordinary `false` default) (flat-shaded — deck.gl only lights the `extruded` shader path — each row also carrying `surface_type`: RoofSurface/WallSurface/GroundSurface, and `fill_color`: a ninja-viewer-style default color per surface_type/cityobject_type, verified against cityjson-threejs-loader's own default palette — `get-fill-color` on `SolidPolygonLayer` reads it automatically when left unauthored, no color attribute required, and an authored `get-fill-color` still overrides it); derived properties (both modes) `roof_height` (area-weighted mean roof height above ground), `eaves_height`, `ridge_height`, `ground_height`, `roof_area`, `surface_count`, `lod`, `cityobject_id`, `cityobject_type`, `parent_id` win over same-named source attributes, plus surfaces-mode-only `polygon`/`outline`/`surface_type`/`fill_color` (`outline` is the face's outer ring flattened and closed — bind a companion `type="PathLayer" get-path="$outline"` layer to it for visible face edges, since surfaces mode is flat-shaded and `SolidPolygonLayer`'s own `wireframe` prop is a no-op when unextruded — always pair one, matching `filter-field`/`filter-range` to the fill layer), and a parent Building's attributes are inherited by its BuildingPart rows; national grids NL/CH/DE/JP/AT/SG reproject automatically including axis order, other EPSG codes fail with an error naming the code; highest LoD wins, pin one with `?om-lod=1.2` (combine as `?om-lod=1.2&om-surfaces=1`, cached independently); `.city.jsonl` fills in as it downloads in either mode — see docs/3d-assets.md), or Arrow IPC `.arrow`/`.feather` — large point datasets stay columnar, GeoArrow line/polygon geometry becomes GeoJSON features, zstd-compressed IPC is handled; other formats plug in via `OmMap.registerFormat({match, parse})`; data URLs accept any scheme the runtime's fetch supports — desktop webviews (Tauri, Electron) pass asset-protocol URLs straight in), inline `<script type="application/json">` (row arrays or column-oriented `{"columns": {"lon": [...], "lat": [...]}}`; must be a DIRECT child of the `<om-layer>`, and when present it wins — omit the `data` attribute), or `wss://` streaming (`key="mmsi"` upserts entities in place, `flush="250ms"` coalesces bursts, `source="name"` selects a `OmMap.registerSource` decoder plugin), or a polled REST snapshot (`refresh="5s"` re-fetches and replaces — for live endpoints that return the full current state). TILED layers: a `{z}/{x}/{y}` `data` template is deck's tile URL for `TileLayer`/`MVTLayer` (NOT rows) — passed through to deck verbatim, never fetched/parsed, so `<om-layer type="TileLayer" data="…/{z}/{x}/{y}.png">` works (raster gets a built-in BitmapLayer sublayer) and `type="MVTLayer" data="…/{z}/{x}/{y}.pbf"` self-renders vector tiles with `get-*` accessors applying to each decoded feature's properties; a tiled layer has no local rows so `ctx.data`/`ctx.stats`/`filter-*` don't apply. Authenticated endpoints: call `OmMap.configureData({ headers: {...} })` in a script — never put tokens in attributes. `$field` accessors work identically on all of them — never write column-index code yourself. One columnar restriction: the `js` full-JS opt-in is not allowed on Arrow/columnar layers (validation will tell you; use `$field` accessors instead). For 3D models use `type="ScenegraphLayer"` with `scenegraph="./model.glb"` (required) and `get-orientation="[0, $heading, 90]"` — the roll of 90 stands Y-up glTF models upright; see docs/3d-assets.md. 3D Tiles use `type="Tile3DLayer"` with `tileset="…/tileset.json"` (NOT `data`). Any pickable layer can opt into deck's DEPTH-pick pass with `pickable="3d"` instead of a bare `pickable` (issue #34) — the resolved click/hover coordinate then carries a real elevation (`{{z}}` in overlay/tooltip templates, `ctx.selection.coordinate`) instead of the ray∩ground-plane guess, so a click on a building face lands ON the face rather than on the ground behind it; `terrain` sets this on itself. For BIM/photogrammetry, `pickable` alone picks a whole TILE — add `pick-features` to pick an individual ELEMENT (a wall, a window, one IFC product), and the `selection` then carries that element's `featureId`/`properties`/`class` from the tile's own `EXT_mesh_features` + `EXT_structural_metadata` (`feature-id-property` selects the ID set, default `_FEATURE_ID_0`). `feature-styles` recolours/fades/highlights by feature ID — an array indexed BY id of `{color: [r,g,b], strength: 0-1, opacity: 0-1}`, live-settable, uploaded as a small lookup texture (no refetch). Use `strength` below 1 to tint rather than replace, or the model's own texture is hidden. Isolate/hide/ghost are DECLARATIVE and mirror the vector `filter-field`/`filter-categories` pair — `feature-filter-field="component"` names the metadata field, then `isolate-features='["Clock"]'` / `hide-features='["Windows","Skylight"]'` / `ghost-features='["Wall"]'` take JSON value lists matched against the tile's property table (`ghost-opacity` tunes the fade, default 0.18). `isolate-features` is EXCLUSIVE (anything unlisted is hidden); hiding is a shader discard so a hidden element also stops being pickable and whatever is behind it becomes selectable. They compose ONTO `feature-styles` (style table supplies colour, these supply visibility), and being attributes they are undoable and story-steppable — prefer them over computing a style table in page JS. Multi-material/multi-primitive models fan out correctly (glTF allows one material per primitive, so real IFC exports are usually many primitives); only genuinely instanced i3dm tiles stay tile-granularity. Limits to state rather than discover: texture-backed IDs — how photogrammetry classification ships — require `load-options='{"gltf":{"loadBuffers":true,"loadImages":true},"image":{"type":"data"}}'`, and omitting `image.type` makes the tileset take MINUTES to appear (loaders.gl otherwise reads the whole ID texture back through a canvas once per vertex); and `opacity` below 1 currently blanks the model. BIM WIDGETS (all need a `pick-features` layer): `<om-widget type="ifc-browser" layer="clinic" fields="ifcClass material container spatialPath" scale-fields="netVolume" rows="7">` groups the model by a property-table field, counts each value, and gives every row I/H/G buttons that WRITE `isolate-features`/`hide-features`/`ghost-features` (so it is a UI over the attributes — undoable and story-steppable; I is MULTI-SELECT, isolating the union of every pressed row or tree node, since `isolate-features` is a list and the layer matches it as a set) while keeping the companion outline layer's `filter-categories` in step; that same select also offers whichever TREES the file supports — Spatial (`spatialPath`), Type (`typePath`), System (`systemPath`) and Classification (`classificationPath`, built by walking `ReferencedSource`) — each expandable with counts aggregated upward and the same I/H/G on every node, so isolating a storey or a system or a CCS code reaches every element under it. Spatial is NOT privileged (on a real Danish project the classification tree covered 3,415 elements to spatial's 660). A tree is not a separate mode, just a group-by on a hierarchy column — LIST it in `fields` to offer it, or set `field="spatialPath"` to open on it — and `loadIfc` emits a hierarchy column ONLY when the file populates it, so a tree that would render empty is never offered; trees appear automatically when available (no need to list them in `fields`) and `field="spatialPath"` opens straight onto one. ONE browser per layer: `feature-filter-field` and the isolate/hide/ghost attributes are single-valued, so two instances on one layer clobber each other. The widget was renamed from `ifc-legend` (still registered as a deprecated alias) because it is a model browser, not a legend. non-physical classes (IfcSpace, IfcOpeningElement) are hidden unless `show-non-physical`, and `no-color` removes the colour select. `<om-widget type="feature-inspector" fields="ifcClass material container netVolume">` (renamed from `ifc-inspector`, kept as an alias) shows the picked element's properties. `<om-widget type="ifc-loader" layer="ifc" federate>` is a drop zone that parses `.ifc` files IN THE BROWSER and builds both layers itself; `federate` accepts SEVERAL models into one co-registered scene (one drop zone, one layer per model, each with a visibility toggle) rather than one widget per discipline, matching how coordination tools append models; a model that turns out to be georeferenced is AUTO-PLACED (its own coordinates/heading/scale written onto the layers the WIDGET created, camera flown there) — but the widget NEVER writes `<om-map>`'s own scene attributes (`basemap`, `terrain`): those are author-owned, and a georeferenced model on a map with neither raises a structured "no spatial context" warning instead of switching one on. COLOUR BY PROPERTY instead of hand-computing a `feature-styles` table: `feature-color-by` (categorical, `feature-palette` overrides the built-in cycle) or `feature-color-scale` (graduated over a numeric field), with `feature-color-strength` (default 0.85) controlling how hard the colour mixes over the model's own material. Setting NEITHER is the default and is meaningful — the model renders in its own IFC surface colours. A graduated ramp needs the field populated: Revit IFC2x3 exports often carry no `IfcElementQuantity`, so every `netVolume` is 0 and the ramp is flat. Widget scripts read the decoded property table with `ctx.features(layerId)` (undefined until the first tile carrying one lands) and re-render on the `features` watch token. `<om-widget type="ifc-clash" layers="arch mep" tolerance="0">` is the CLASH OVERLAY over two co-registered model layers: it flags element pairs whose bounding boxes interpenetrate, colours both sides via `feature-styles`, and flies to the centre of each overlap. v1 is an axis-aligned box test — fast and serverless, but it over-reports anything diagonal and says nothing about which clashes matter; zones/spaces/openings/proxies and same-class-same-name pairs are excluded as noise. The header carries an overlay on/off switch; an isolation mode select (None/Dim/Hide) sits above the results list and applies once a row is selected (with nothing selected, nothing is hidden), and clicking a row FOCUSES that clash (chosen pair at full strength, every other clashing element dropped to a faint tint, camera flown to the overlap centre) — without that, everything is highlighted and nothing is. Results are GROUPED by the side-A element with a count (one wall crossing four ducts is one row), and the two model selects appear only when more than two models are loaded. It CHECKS co-registration (matching `site-origin`) and says so when it fails, because two mis-registered models report zero clashes exactly like two clean ones. Persisting/sharing results is BCF's job and is out of scope. `<om-layer type="BIMLayer" src="./model.ifc">` is the declarative counterpart to `ifc-loader`/`loadIfc`: point it at a BIM source file (an .ifc today) and it runs the loader itself the moment src resolves — no pre-baked tileset, no site-origin/site-heading/site-scale (the file's own georeference is read and applied automatically; not wired up yet: an authored site-origin on a BIMLayer does not override it), and no separate PathLayer for the outline overlay (added automatically). `pick-features` defaults ON (unlike a plain Tile3DLayer); feature-filter-field/feature-styles/isolate-hide-ghost/feature-color-by/ghost-opacity all work unchanged, since BIMLayer forwards them to a real Tile3DLayer it builds internally. Known gap: the outline overlay does not yet follow isolate/hide/ghost the way the mesh does. Reach for BIMLayer when the model is fixed and known ahead of time; reach for ifc-loader when a visitor picks the file or several models need to federate. IN-BROWSER IFC: `loadIfc(bytes)` parses an `.ifc` with web-ifc (WASM, MPL-2.0, CDN-fetched on first use — NOT a package dependency; `configureIfc({wasmPath})` self-hosts) and returns `{tilesetUrl, edgesUrl, loadOptions, features (rows carry `ifcClass`/`name`/`material`/`container`/`netVolume`/`spatialPath`/`typePath`/`systemPath`/`classificationPath` — hierarchy columns joined by U+001F that the model trees navigate, each emitted only when the file populates it), lonLat, georeferenced, heading, scale, originSource, headingSource, stats, timings, bounds, revoke()}`; the output IS a tileset so picking/styling/`site-*`/isolate-hide-ghost work unchanged. CALL `revoke()` when swapping models — blob URLs are held by the document. FEDERATION: pass the first model's returned `origin` as `LoadIfcOptions.origin` for every later model of the same building, or each is centred on its own bounding box and they drift apart — a clash pass then finds nothing, which looks identical to a clean model. `ifc-loader` shares one origin AND one placement (`site-origin`/`site-heading`/`site-scale`) per map automatically — discipline files routinely declare IfcSite coordinates kilometres apart for the same building, so the first model loaded decides where it goes and the rest follow (`independent` opts out). Every element also carries a tile-local bounding box in the property table (`bboxMinE`/`bboxMinN`/`bboxMinU`/`bboxMaxE`/`bboxMaxN`/`bboxMaxU`). POSITION is read from the file preferring the trustworthy route: `IfcMapConversion` (a surveyed placement into a named projected CRS) WINS over `IfcSite.RefLatitude`/`RefLongitude`, which is very often an authoring default; `originSource` reports which was used, and anything other than `"map-conversion"` raises a structured `"warning"` through the same validation channel other `om-layer` errors use (the on-page panel with `validate` set, `om-validation-error`'s `detail.warnings`) for both `BIMLayer` and `ifc-loader`, once per layer — it never flips `valid` false, only flags the position may be off by tens of metres with no rotation correction applied; override with `site-origin`/`site-heading` or a proper `IfcMapConversion`. Un-projecting a map conversion is supported for WGS84 UTM zones (EPSG:326xx/327xx) and DECLINED with a warning for anything else — a guessed projection lands the model in another country while looking plausible. Eastings/northings are in the target CRS unit, frequently MILLIMETRES. Reading a file correctly and the model being somewhere sensible are SEPARATE problems: all three prepared samples declare placeholders (clinic on Revit's Boston default, which is a 1630 graveyard; duplex on a Chicago city-centre point; bridge — which does carry a real IfcMapConversion — into the mid-Pacific). `headingSource` distinguishes "map-conversion"/"true-north" (read) from "assumed" (file was silent) — report the assumption, never let it read as a measurement. `<om-map>` reads camera attributes ONCE at init, so `setAttribute("center", …)` after mount moves nothing: use `map.flyTo(lonLat, zoom)`. GEOREFERENCING is declarative on `Tile3DLayer` and `PathLayer`: `site-origin="[lng, lat]"` (or `[lng, lat, elevation]`) OVERRIDES the position baked into a tileset's root transform, `site-heading` is a bearing in degrees CLOCKWISE from true north (on its own it rotates the model where it stands), `site-scale` is a uniform multiplier; rotation and scale pivot on the model's own anchor, not the tileset origin. An IFC model is a PAIR of layers — the mesh tileset plus a `PathLayer` outline overlay whose paths are local east/north/up METRES — and both need the same three values or the building separates from its own edges. Never trust a model's declared position without looking at it: authoring tools ship a default project location that is indistinguishable from a survey (the buildingSMART Medical-Dental Clinic sample carries Revit's Boston default, the Duplex a Chicago city-centre point, so both land on occupied downtown blocks at an arbitrary rotation), `IfcMapConversion` is absent from most IFC2x3 exports, and `TrueNorth` is routinely unset. Editing `site-*` on a live Tile3DLayer reloads the tileset (deck.gl only reloads on a URL change); the PathLayer updates as a uniform. GeoTIFF/COG rasters use `type="COGLayer"` with `src="./dem.tif"` (NOT `data` — rasters stream tiles by Range request, they are not parsed rows). COG v2 surface (0.7.0): `bands` selects 1-based bands — one (`bands="4"`, colormap-eligible) or an `[r,g,b]` triple (`bands="[8,4,3]"`, false-color composite); band switches NEVER refetch (the decode is retained; textures re-upload from memory). `min`/`max` set the rescale window — a number broadcasts, a triple pairs per-band with a `bands` triple; `rescale="auto"` reads the window from GDAL statistics tags (coarsest-overview sample fallback, with a warning that the window is data-derived); non-8-bit sources with NO styling attrs auto-stretch automatically now (the "my COG is blank" fix — a dev notice reports the derived window; author min/max for stable styling). `colormap` picks a bundled ramp for single-band sources (gray, viridis, plasma, inferno, magma, cividis, rdylgn, rdbu, spectral, terrain, jet, turbo, ylorrd), `reverse` flips it, `stretch="log|sqrt"` and `gamma="1.8"` shape the display curve, `nodata` overrides the source sentinel (renders transparent). Plain 8-bit RGB COGs need no styling attributes; paletted GeoTIFFs render via their embedded color table and emit a classes legend when few palette entries are used. Every styling change is a GPU uniform (no refetch); the legend derives automatically (authored or auto window). PIXEL IDENTIFY: clicks/hovers over a COGLayer with no vector feature carry the pixel's per-band values through the ordinary selection — overlays/tooltips interpolate `{{value}}`, `{{values}}`, `{{band_1}}`…; `identify="off"` opts out (drops the CPU-retained decode — recommended for memory-tight WebViews; band switches on that layer then refetch tiles instead of re-uploading from memory). Source failures (CORS-refused Range requests, non-COG plain GeoTIFFs, out-of-range bands) surface as structured validation errors with fix strings instead of a silent blank layer. Sources should be Cloud-Optimized (`gdal_translate -of COG`; plain GeoTIFFs render with a warning, re-reading the full image per view).
+- `<om-layer id="..." type="ScatterplotLayer" data="./points.json">` — any deck.gl layer class by `type` (39 layer types total: 32 bundled deck.gl core/geo/aggregation/mesh layers, plus the native `COGLayer`/`ZarrLayer` raster types, `ImageOverlay` for georeferenced drone JPEGs, `BIMLayer` for BIM source files, and `Route`/`Tracking` for routing/live-tracking — see the dedicated bullet below), plus `PopupLayer` (WebGL badges/labels at scale: `layout="badge|pin-label|card"`, `min-zoom`/`max-zoom`). Every type's full attribute list ships in the package's `onlymapjs.html-data.json` — consult it instead of guessing attribute names or reading the minified dist. TextLayer's default font atlas covers ASCII only: set `character-set` when label text carries other glyphs (`—`, `·`, accents) or deck warns and renders them blank. External layer classes register via `OmMap.registerLayer({type, deckClass, props})` — build them on `@nika-js/onlymap/deck` (the bundled `CompositeLayer`/`TileLayer`/… re-exports), never a separately-installed deck.gl (different class hierarchy, breaks in the renderer); function-valued props ride the subclass's `static defaultProps`; register at module top level BEFORE the manifest mounts (see docs/custom-layers.md). Data: `data` URL (JSON, GeoJSON, CSV/TSV `.csv` — parsed to typed columns — Shapefile `.shp` (+`.dbf` attributes) and KML `.kml` as GeoJSON features, GPX `.gpx` (waypoints/tracks/routes → features tagged `_gpxKind`; a `#waypoints`/`#tracks`/`#routes` URL fragment selects one part), FlatGeobuf `.fgb` (cloud-native binary vector, whole-file decode), GeoParquet `.parquet`/`.geoparquet` (cloud-native columnar vector — all-Point files stay columnar like Arrow, lines/polygons become GeoJSON features; requires the file's `geo` metadata with WKB geometry, and CRS84/EPSG:4326 — a projected CRS is a loud error telling you to reproject, not a silent misplacement; snappy/gzip/zstd row-group compression handled), CityJSON `.city.json` / CityJSONSeq `.city.jsonl` (semantic 3D city models — 3DBAG, PLATEAU — decoded to one of two shapes by the `data` URL, no CityJSON layer type: default → extruded footprints, `type="GeoJsonLayer" extruded get-elevation="$roof_height"`; `?om-surfaces=1` → one row PER FACE at its own real per-vertex height so a pitched LoD2.2 roof actually looks pitched, `type="SolidPolygonLayer" get-polygon="$polygon" full3d` (`extruded` stays at its ordinary `false` default) (flat-shaded — deck.gl only lights the `extruded` shader path — each row also carrying `surface_type`: RoofSurface/WallSurface/GroundSurface, and `fill_color`: a ninja-viewer-style default color per surface_type/cityobject_type, verified against cityjson-threejs-loader's own default palette — `get-fill-color` on `SolidPolygonLayer` reads it automatically when left unauthored, no color attribute required, and an authored `get-fill-color` still overrides it); derived properties (both modes) `roof_height` (area-weighted mean roof height above ground), `eaves_height`, `ridge_height`, `ground_height`, `roof_area`, `surface_count`, `lod`, `cityobject_id`, `cityobject_type`, `parent_id` win over same-named source attributes, plus surfaces-mode-only `polygon`/`outline`/`surface_type`/`fill_color` (`outline` is the face's outer ring flattened and closed — bind a companion `type="PathLayer" get-path="$outline"` layer to it for visible face edges, since surfaces mode is flat-shaded and `SolidPolygonLayer`'s own `wireframe` prop is a no-op when unextruded — always pair one, matching `filter-field`/`filter-range` to the fill layer), and a parent Building's attributes are inherited by its BuildingPart rows; national grids NL/CH/DE/JP/AT/SG reproject automatically including axis order, other EPSG codes fail with an error naming the code; highest LoD wins, pin one with `?om-lod=1.2` (combine as `?om-lod=1.2&om-surfaces=1`, cached independently); `.city.jsonl` fills in as it downloads in either mode — see docs/3d-assets.md), or Arrow IPC `.arrow`/`.feather` — large point datasets stay columnar, GeoArrow line/polygon geometry becomes GeoJSON features, zstd-compressed IPC is handled; other formats plug in via `OmMap.registerFormat({match, parse})`; data URLs accept any scheme the runtime's fetch supports — desktop webviews (Tauri, Electron) pass asset-protocol URLs straight in), inline `<script type="application/json">` (row arrays or column-oriented `{"columns": {"lon": [...], "lat": [...]}}`; must be a DIRECT child of the `<om-layer>`, and when present it wins — omit the `data` attribute), or `wss://` streaming (`key="mmsi"` upserts entities in place, `flush="250ms"` coalesces bursts, `source="name"` selects a `OmMap.registerSource` decoder plugin), or a polled REST snapshot (`refresh="5s"` re-fetches and replaces — for live endpoints that return the full current state). TILED layers: a `{z}/{x}/{y}` `data` template is deck's tile URL for `TileLayer`/`MVTLayer` (NOT rows) — passed through to deck verbatim, never fetched/parsed, so `<om-layer type="TileLayer" data="…/{z}/{x}/{y}.png">` works (raster gets a built-in BitmapLayer sublayer) and `type="MVTLayer" data="…/{z}/{x}/{y}.pbf"` self-renders vector tiles with `get-*` accessors applying to each decoded feature's properties; a tiled layer has no local rows so `ctx.data`/`ctx.stats`/`filter-*` don't apply. Authenticated endpoints: call `OmMap.configureData({ headers: {...} })` in a script — never put tokens in attributes. `$field` accessors work identically on all of them — never write column-index code yourself. One columnar restriction: the `js` full-JS opt-in is not allowed on Arrow/columnar layers (validation will tell you; use `$field` accessors instead). For 3D models use `type="ScenegraphLayer"` with `scenegraph="./model.glb"` (required) and `get-orientation="[0, $heading, 90]"` — the roll of 90 stands Y-up glTF models upright; `color`/`get-color` tint the model as a shader multiply over its own baked texture (default opaque white = no tint; `get-color` is `$field`-driven like other `*Color` accessors, so a fleet recolors by data without swapping model files); see docs/3d-assets.md. 3D Tiles use `type="Tile3DLayer"` with `tileset="…/tileset.json"` (NOT `data`). Any pickable layer can opt into deck's DEPTH-pick pass with `pickable="3d"` instead of a bare `pickable` (issue #34) — the resolved click/hover coordinate then carries a real elevation (`{{z}}` in overlay/tooltip templates, `ctx.selection.coordinate`) instead of the ray∩ground-plane guess, so a click on a building face lands ON the face rather than on the ground behind it; `terrain` sets this on itself. For BIM/photogrammetry, `pickable` alone picks a whole TILE — add `pick-features` to pick an individual ELEMENT (a wall, a window, one IFC product), and the `selection` then carries that element's `featureId`/`properties`/`class` from the tile's own `EXT_mesh_features` + `EXT_structural_metadata` (`feature-id-property` selects the ID set, default `_FEATURE_ID_0`). `feature-styles` recolours/fades/highlights by feature ID — an array indexed BY id of `{color: [r,g,b], strength: 0-1, opacity: 0-1}`, live-settable, uploaded as a small lookup texture (no refetch). Use `strength` below 1 to tint rather than replace, or the model's own texture is hidden. Isolate/hide/ghost are DECLARATIVE and mirror the vector `filter-field`/`filter-categories` pair — `feature-filter-field="component"` names the metadata field, then `isolate-features='["Clock"]'` / `hide-features='["Windows","Skylight"]'` / `ghost-features='["Wall"]'` take JSON value lists matched against the tile's property table (`ghost-opacity` tunes the fade, default 0.18). `isolate-features` is EXCLUSIVE (anything unlisted is hidden); hiding is a shader discard so a hidden element also stops being pickable and whatever is behind it becomes selectable. They compose ONTO `feature-styles` (style table supplies colour, these supply visibility), and being attributes they are undoable and story-steppable — prefer them over computing a style table in page JS. Multi-material/multi-primitive models fan out correctly (glTF allows one material per primitive, so real IFC exports are usually many primitives); only genuinely instanced i3dm tiles stay tile-granularity. Limits to state rather than discover: texture-backed IDs — how photogrammetry classification ships — require `load-options='{"gltf":{"loadBuffers":true,"loadImages":true},"image":{"type":"data"}}'`, and omitting `image.type` makes the tileset take MINUTES to appear (loaders.gl otherwise reads the whole ID texture back through a canvas once per vertex); and `opacity` below 1 currently blanks the model. BIM WIDGETS (all need a `pick-features` layer): `<om-widget type="ifc-browser" layer="clinic" fields="ifcClass material container spatialPath" scale-fields="netVolume" rows="7">` groups the model by a property-table field, counts each value, and gives every row I/H/G buttons that WRITE `isolate-features`/`hide-features`/`ghost-features` (so it is a UI over the attributes — undoable and story-steppable; I is MULTI-SELECT, isolating the union of every pressed row or tree node, since `isolate-features` is a list and the layer matches it as a set) while keeping the companion outline layer's `filter-categories` in step; that same select also offers whichever TREES the file supports — Spatial (`spatialPath`), Type (`typePath`), System (`systemPath`) and Classification (`classificationPath`, built by walking `ReferencedSource`) — each expandable with counts aggregated upward and the same I/H/G on every node, so isolating a storey or a system or a CCS code reaches every element under it. Spatial is NOT privileged (on a real Danish project the classification tree covered 3,415 elements to spatial's 660). A tree is not a separate mode, just a group-by on a hierarchy column — LIST it in `fields` to offer it, or set `field="spatialPath"` to open on it — and `loadIfc` emits a hierarchy column ONLY when the file populates it, so a tree that would render empty is never offered; trees appear automatically when available (no need to list them in `fields`) and `field="spatialPath"` opens straight onto one. ONE browser per layer: `feature-filter-field` and the isolate/hide/ghost attributes are single-valued, so two instances on one layer clobber each other. The widget was renamed from `ifc-legend` (still registered as a deprecated alias) because it is a model browser, not a legend. non-physical classes (IfcSpace, IfcOpeningElement) are hidden unless `show-non-physical`, and `no-color` removes the colour select. `<om-widget type="feature-inspector" fields="ifcClass material container netVolume">` (renamed from `ifc-inspector`, kept as an alias) shows the picked element's properties. `<om-widget type="ifc-loader" layer="ifc" federate>` is a drop zone that parses `.ifc` files IN THE BROWSER and builds both layers itself; `federate` accepts SEVERAL models into one co-registered scene (one drop zone, one layer per model, each with a visibility toggle) rather than one widget per discipline, matching how coordination tools append models; a model that turns out to be georeferenced is AUTO-PLACED (its own coordinates/heading/scale written onto the layers the WIDGET created, camera flown there) — but the widget NEVER writes `<om-map>`'s own scene attributes (`basemap`, `terrain`): those are author-owned, and a georeferenced model on a map with neither raises a structured "no spatial context" warning instead of switching one on. COLOUR BY PROPERTY instead of hand-computing a `feature-styles` table: `feature-color-by` (categorical, `feature-palette` overrides the built-in cycle) or `feature-color-scale` (graduated over a numeric field), with `feature-color-strength` (default 0.85) controlling how hard the colour mixes over the model's own material. Setting NEITHER is the default and is meaningful — the model renders in its own IFC surface colours. A graduated ramp needs the field populated: Revit IFC2x3 exports often carry no `IfcElementQuantity`, so every `netVolume` is 0 and the ramp is flat. Widget scripts read the decoded property table with `ctx.features(layerId)` (undefined until the first tile carrying one lands) and re-render on the `features` watch token. `<om-widget type="ifc-clash" layers="arch mep" tolerance="0">` is the CLASH OVERLAY over two co-registered model layers: it flags element pairs whose bounding boxes interpenetrate, colours both sides via `feature-styles`, and flies to the centre of each overlap. v1 is an axis-aligned box test — fast and serverless, but it over-reports anything diagonal and says nothing about which clashes matter; zones/spaces/openings/proxies and same-class-same-name pairs are excluded as noise. The header carries an overlay on/off switch; an isolation mode select (None/Dim/Hide) sits above the results list and applies once a row is selected (with nothing selected, nothing is hidden), and clicking a row FOCUSES that clash (chosen pair at full strength, every other clashing element dropped to a faint tint, camera flown to the overlap centre) — without that, everything is highlighted and nothing is. Results are GROUPED by the side-A element with a count (one wall crossing four ducts is one row), and the two model selects appear only when more than two models are loaded. It CHECKS co-registration (matching `site-origin`) and says so when it fails, because two mis-registered models report zero clashes exactly like two clean ones. Persisting/sharing results is BCF's job and is out of scope. `<om-layer type="BIMLayer" src="./model.ifc">` is the declarative counterpart to `ifc-loader`/`loadIfc`: point it at a BIM source file (an .ifc today) and it runs the loader itself the moment src resolves — no pre-baked tileset, no site-origin/site-heading/site-scale (the file's own georeference is read and applied automatically; not wired up yet: an authored site-origin on a BIMLayer does not override it), and no separate PathLayer for the outline overlay (added automatically). `pick-features` defaults ON (unlike a plain Tile3DLayer); feature-filter-field/feature-styles/isolate-hide-ghost/feature-color-by/ghost-opacity all work unchanged, since BIMLayer forwards them to a real Tile3DLayer it builds internally. Known gap: the outline overlay does not yet follow isolate/hide/ghost the way the mesh does. Reach for BIMLayer when the model is fixed and known ahead of time; reach for ifc-loader when a visitor picks the file or several models need to federate. IN-BROWSER IFC: `loadIfc(bytes)` parses an `.ifc` with web-ifc (WASM, MPL-2.0, CDN-fetched on first use — NOT a package dependency; `configureIfc({wasmPath})` self-hosts) and returns `{tilesetUrl, edgesUrl, loadOptions, features (rows carry `ifcClass`/`name`/`material`/`container`/`netVolume`/`spatialPath`/`typePath`/`systemPath`/`classificationPath` — hierarchy columns joined by U+001F that the model trees navigate, each emitted only when the file populates it), lonLat, georeferenced, heading, scale, originSource, headingSource, stats, timings, bounds, revoke()}`; the output IS a tileset so picking/styling/`site-*`/isolate-hide-ghost work unchanged. CALL `revoke()` when swapping models — blob URLs are held by the document. FEDERATION: pass the first model's returned `origin` as `LoadIfcOptions.origin` for every later model of the same building, or each is centred on its own bounding box and they drift apart — a clash pass then finds nothing, which looks identical to a clean model. `ifc-loader` shares one origin AND one placement (`site-origin`/`site-heading`/`site-scale`) per map automatically — discipline files routinely declare IfcSite coordinates kilometres apart for the same building, so the first model loaded decides where it goes and the rest follow (`independent` opts out). Every element also carries a tile-local bounding box in the property table (`bboxMinE`/`bboxMinN`/`bboxMinU`/`bboxMaxE`/`bboxMaxN`/`bboxMaxU`). POSITION is read from the file preferring the trustworthy route: `IfcMapConversion` (a surveyed placement into a named projected CRS) WINS over `IfcSite.RefLatitude`/`RefLongitude`, which is very often an authoring default; `originSource` reports which was used, and anything other than `"map-conversion"` raises a structured `"warning"` through the same validation channel other `om-layer` errors use (the on-page panel with `validate` set, `om-validation-error`'s `detail.warnings`) for both `BIMLayer` and `ifc-loader`, once per layer — it never flips `valid` false, only flags the position may be off by tens of metres with no rotation correction applied; override with `site-origin`/`site-heading` or a proper `IfcMapConversion`. Un-projecting a map conversion is supported for WGS84 UTM zones (EPSG:326xx/327xx) and DECLINED with a warning for anything else — a guessed projection lands the model in another country while looking plausible. Eastings/northings are in the target CRS unit, frequently MILLIMETRES. Reading a file correctly and the model being somewhere sensible are SEPARATE problems: all three prepared samples declare placeholders (clinic on Revit's Boston default, which is a 1630 graveyard; duplex on a Chicago city-centre point; bridge — which does carry a real IfcMapConversion — into the mid-Pacific). `headingSource` distinguishes "map-conversion"/"true-north" (read) from "assumed" (file was silent) — report the assumption, never let it read as a measurement. `<om-map>` reads camera attributes ONCE at init, so `setAttribute("center", …)` after mount moves nothing: use `map.flyTo(lonLat, zoom)`. GEOREFERENCING is declarative on `Tile3DLayer` and `PathLayer`: `site-origin="[lng, lat]"` (or `[lng, lat, elevation]`) OVERRIDES the position baked into a tileset's root transform, `site-heading` is a bearing in degrees CLOCKWISE from true north (on its own it rotates the model where it stands), `site-scale` is a uniform multiplier; rotation and scale pivot on the model's own anchor, not the tileset origin. An IFC model is a PAIR of layers — the mesh tileset plus a `PathLayer` outline overlay whose paths are local east/north/up METRES — and both need the same three values or the building separates from its own edges. Never trust a model's declared position without looking at it: authoring tools ship a default project location that is indistinguishable from a survey (the buildingSMART Medical-Dental Clinic sample carries Revit's Boston default, the Duplex a Chicago city-centre point, so both land on occupied downtown blocks at an arbitrary rotation), `IfcMapConversion` is absent from most IFC2x3 exports, and `TrueNorth` is routinely unset. Editing `site-*` on a live Tile3DLayer reloads the tileset (deck.gl only reloads on a URL change); the PathLayer updates as a uniform. GeoTIFF/COG rasters use `type="COGLayer"` with `src="./dem.tif"` (NOT `data` — rasters stream tiles by Range request, they are not parsed rows). COG v2 surface (0.7.0): `bands` selects 1-based bands — one (`bands="4"`, colormap-eligible) or an `[r,g,b]` triple (`bands="[8,4,3]"`, false-color composite); band switches NEVER refetch (the decode is retained; textures re-upload from memory). `min`/`max` set the rescale window — a number broadcasts, a triple pairs per-band with a `bands` triple; `rescale="auto"` reads the window from GDAL statistics tags (coarsest-overview sample fallback, with a warning that the window is data-derived); non-8-bit sources with NO styling attrs auto-stretch automatically now (the "my COG is blank" fix — a dev notice reports the derived window; author min/max for stable styling). `colormap` picks a bundled ramp for single-band sources (gray, viridis, plasma, inferno, magma, cividis, rdylgn, rdbu, spectral, terrain, jet, turbo, ylorrd), `reverse` flips it, `stretch="log|sqrt"` and `gamma="1.8"` shape the display curve, `nodata` overrides the source sentinel (renders transparent). Plain 8-bit RGB COGs need no styling attributes; paletted GeoTIFFs render via their embedded color table and emit a classes legend when few palette entries are used. Every styling change is a GPU uniform (no refetch); the legend derives automatically (authored or auto window). PIXEL IDENTIFY: clicks/hovers over a COGLayer with no vector feature carry the pixel's per-band values through the ordinary selection — overlays/tooltips interpolate `{{value}}`, `{{values}}`, `{{band_1}}`…; `identify="off"` opts out (drops the CPU-retained decode — recommended for memory-tight WebViews; band switches on that layer then refetch tiles instead of re-uploading from memory). Source failures (CORS-refused Range requests, non-COG plain GeoTIFFs, out-of-range bands) surface as structured validation errors with fix strings instead of a silent blank layer. Sources should be Cloud-Optimized (`gdal_translate -of COG`; plain GeoTIFFs render with a warning, re-reading the full image per view).
 - Geotagged drone JPEGs are the library-owned `ImageOverlay` type, not row `data` and not a raw `BitmapLayer`: `<om-layer type="ImageOverlay" src="./photo.jpg" georeference="exif">`. It reads GPS/relative altitude/camera/focal length plus DJI gimbal metadata through the configured fetch policy, waits before `ready`, bakes yaw/roll, and computes visualization-grade flat-ground bounds. Unknown cameras need `sensor-width-mm` + `sensor-height-mm` (and `focal-length-mm` when EXIF lacks it). For collaborative/saved maps call `OmMap.resolveImageOverlay(fileOrUrl)`, upload its returned `image`, then reconstruct using `src` + the returned explicit `bounds` (no EXIF fetch). Use `COGLayer` for large orthomosaics; see docs/image-overlays.md.
 - Chunked N-dimensional Zarr / GeoZarr rasters (climate/weather grids, datacubes) are the library-owned `ZarrLayer` type (built on `@developmentseed/deck.gl-zarr` + zarrita, a lazy chunk): `<om-layer type="ZarrLayer" src="./x.zarr" variable="temp" select="time=0" colormap="viridis" min="…" max="…">`. `src` not `data` (chunks stream through the layer's reader, never parsed rows). Pick the `variable` and pin EVERY non-spatial dimension in `select` ("init_time=0, lead_time=0, ensemble_member=0"); the two spatial dims are handled for you (a 2-D array needs no select). A GeoZarr-compliant store georeferences itself; a plain Zarr needs manual `bounds="[w,s,e,n]"` + `crs="EPSG:4326"` + `spatial-dims="<yName> <xName>"` (bounds without crs+spatial-dims is a validation error). `min`/`max`/`colormap`/`nodata` and the auto legend reuse the exact COGLayer raster pipeline. Beware store chunking: a dataset chunked coarsely over non-spatial dims (e.g. all forecast steps in one chunk) decodes far more than the pinned frame needs. `src` may be any absolute URL (`https://…/store.zarr`) — an external/remote store works with no server setup (a static host serves Zarr's extensionless chunk keys natively), but zarrita fetches it directly from the browser so the store MUST send CORS headers (`Access-Control-Allow-Origin`), and it must be PUBLIC — authenticated stores are not yet supported (ZarrLayer uses zarrita's own fetch, not `OmMap.configureData`).
 - `<om-widget type="legend|layer-switcher|basemap-switcher|lighting|clip-box|zoom-controls|undo-redo|scale-bar|attribution|filter|vega-lite|dynamic-chart|measure" position="bottom-end">` — static UI panels. `position` takes one of 8 managed slots (logical, RTL-aware: `top-start|top-center|top-end|center-start|center-end|bottom-start|bottom-center|bottom-end`; legacy corners `top-left` etc. alias) — same-slot widgets stack with flush edges and a shared gap (never overlap); provider attribution joins `bottom-end` and the license badge joins `bottom-start` as in-flow members, so required chrome never covers a widget; `order="1"` orders within a slot; adjacent compact button widgets (zoom-controls, undo-redo, widgets-toggle) auto-merge into ONE control group with dividers (`cluster="false"` opts a widget out); `position="manual"` renders a plain block you place with your own CSS (even outside the map). At narrow map widths managed widgets auto-fold into one accessible drawer per map side — at ≤640px when some row (top/center/bottom) has widgets on more than one of its slots, and only at ≤416px when every occupied row is one-sided (a lone legend at center-start keeps the row to itself and survives narrower maps); `fold="never"` exempts an essential widget, `widgets-fold="off"` opts the map out, `--om-widget-fold-breakpoint` changes the threshold. Layout tokens on `<om-map>`: `widget-style="gap:10 opacity:0.9 inset:16"` (keys inset/gap/inset-x/-y/gap-x/-y/opacity/radius/size, px except opacity) or the `--om-widget-inset-x/-y/-gap-x/-gap-y/-opacity/-radius/-fold-breakpoint` custom properties. Built-ins are themeable from page CSS via custom properties (they inherit through the shadow root): `om-map { --om-widget-bg: #111827; --om-widget-fg: #f9fafb; }` — full set: `--om-widget-bg/-fg/-muted/-border/-hover-bg/-accent`; scope to a single widget with an `om-widget[type=legend]` selector instead. `measure` is a geodesic ruler: `modes="distance area volume"` (default `distance area`), `units="metric|imperial|nautical"` — click the map to place points, live per-segment + total labels render on the map, and it dispatches an `om-measure` event; it reuses the draw capture stack (measure/draw mutually exclusive) and its geometry is ephemeral (never saved, not an undo step) — see the dedicated `measure`/`volume` bullet above for the full event shape and the `profile`/`density`/`swell`/`shrink`/`deadband` volume-mode attributes. `scale-bar` now takes the same `units`. `dynamic-chart` is the same Vega-Lite rendering as `vega-lite` but fed by a live DOM event instead of a layer: `on="<event-name>"` (required) + `series-field="<name>"` (default `series`) reads `event.detail[seriesField]` as `data.values` and re-embeds on every event where it's a present array, at a fixed `width`; omitting the field on a later event freezes the chart with no separate pause API — built for a feature (a drawn line's elevation profile) that computes its own series live and has no layer to bind to. No `type` + HTML + `<script type="om/widget">` = custom widget with `ctx` (`ctx.layers`, `ctx.data(id)`, `ctx.dataInViewport(id)`, `ctx.stats(id, field)`, `ctx.viewport`, `ctx.selection`, `ctx.emit(action, payload)`), `this.watch = ['data:<layerId>', 'viewport', 'selection', 'layers', 'history', 'basemap']` (`layers` also fires on visibility/filter changes; `basemap` on basemap switches; `history` on undo/redo availability), `this.$(sel)`, `vegaEmbed`/`d3` as globals.
@@ -43,6 +43,7 @@ Programmatic/native bridge rule: `MapController.setLayers()` accepts normal func
 - `<om-behavior on="click|hover|drag|load|data-loaded" layer="..." action="...">` — declarative interaction. Built-in actions: `show-overlay`, `hide-overlay`, `show-tooltip`, `hide-tooltip`, `toggle-layer`, `set-pickable` (`{layer, pickable: true|false|"3d"}` — runtime picking/popup toggle, story-capturable; the per-layer popup on/off switch a viewer-facing export needs), `filter-layer`, `highlight-feature`, `zoom-to-feature`, `set-basemap`, `undo`, `redo`; scene/tool actions `set-lighting`, `set-terrain`, `set-clip-box` (`{min,max,invert?,highlight?}` / `{clear:true}`), `clip-box-edit` (`{editing}`), `export-region-3d` (`{target?, format?:"glb"|"b3dm"}` — what the draw widget's `export-3d` button emits), the draw actions `draw-mode` (`{target?:"sketch", mode:"point"|"line"|"polygon"|null}` — null exits drawing), `draw-commit`, `draw-cancel`, `draw-delete` (removes the last shape), `draw-clear`, `draw-config` (`{target?, autosave?, fillColor?, lineColor?}`) and `draw-save` (`{target?, format?:"download"|"file-system"|"both"}`) — every one of these drives the same store a `data="draw:<target>"` layer reads, so a custom toolbar can replace `<om-widget type="draw">` entirely; and the measure actions `measure-mode` (`{mode:"distance"|"area"|"volume"|null}`), `measure-units`, `measure-clear`, `measure-config` (`{profile?, baseSurface?, density?, swell?, shrink?, deadband?}`), `measure-flat-target-plane` (`{flat}`). One payload contract everywhere: `{ layer, target, feature, featureId, coordinate }`.
 - Undo/redo is built in: user-facing manifest changes (layer toggles, filter changes, basemap switches, element add/remove, drawn sketches) are recorded automatically — the manifest is the state. `<om-widget type="undo-redo">` renders the buttons; Cmd/Ctrl-Z, Shift-Cmd/Ctrl-Z, and Ctrl-Y work on any map (text inputs keep their native undo). Camera moves, hover effects, and story playback are deliberately NOT undo steps. Widget scripts: `ctx.history.canUndo/canRedo` with watch token `history`; `ctx.emit("undo")`/`ctx.emit("redo")`.
 - `<om-fallback>` — static no-JS fallback, direct child of `<om-map>` (one per map, no attributes, plain HTML content — links allowed). Shown ONLY where scripts never run (chat-app/email file previews — iOS QuickLook renders HTML attachments with JS off — file managers, sandboxed webviews); hidden automatically once the map boots. GOOD PRACTICE: include one on every complete page, especially pages that may be shared as a file. To PRODUCE a shareable single file, `npx @nika-js/onlymap export map.html` embeds every relative data/src/scenegraph URL as a data: URL (format detection preserved — the original extension rides a #name.ext fragment), pins relative library refs to the CDN, adds a generic om-fallback when missing, and warns when embedded data is heavy (>1 MB per file / >5 MB total) or when a source cannot be embedded (shapefile sidecars, COG Range streams, Zarr directories, {z}/{x}/{y} templates — those need hosting) ("This interactive map requires JavaScript — open this file in a web browser", plus a hosted-version link when one exists). Without one, the stylesheet shows a generic text-only banner. The gate is pure CSS (`om-map:not(:defined)` in onlymapjs.css) refined by the `scripting` media feature on 2023+ engines: with scripting DISABLED the fallback shows instantly; with scripting ENABLED it never flashes during a slow load and only surfaces after a ~4s grace with load-failure wording (blocked/unreachable bundle). The CSS must load without JS — a real `<link rel="stylesheet">` or inlined `<style>` on no-build pages; a bundler-emitted stylesheet is fine in npm projects.
+- PRINT LAYOUTS (`@nika-js/onlymap/cartograph`, separate lazy entry — the core bundle does not grow): `<om-cartograph size="A3" orientation="landscape">` is a standalone page measured in MILLIMETRES carrying `<om-frame>`s (live `<om-map>` with the frame's own camera, or `mode="static"` georeferenced rasters placed by `crs`+`corners`) plus cartographic furniture: `<om-legend for="...">` (rows DERIVED from the live map's own symbology incl. classified breaks; author literal `<om-legend-row>`s with `derived="false"`), `<om-scalebar>` (metric|imperial|nautical), `<om-north>`, `<om-graticule>` (4326/3857/UTM), `<om-shape>`, `<om-image>`, `<om-text>` with `{{title}}`/`{{date}}`/`{{frame.<id>.scale}}`/`{{atlas.<field>}}` tokens. `overview-of="<frame-id>"` draws a locator inset's footprint. `<om-atlas for="<frame-id>" layer="<layer-id>">` renders one page per feature (sequentially through ONE live map — WebGL context limits make cloned pages fail at scale) and exports a ZIP. Print via `page.print()` or `?print=1` (true page size via @page; live frames swap to print-DPI captures first), PNG via `renderCartograph(page, {dpi})` or `?export=png&dpi=300`. `bleed`/`crop-marks`/`safe-zone` (screen-only guide)/`flatten` are press-shop attributes; `cvd="deuteranopia|protanopia|tritanopia|achromatopsia"` simulates colour-vision deficiencies page-wide and `lintLegendColours()` flags collapsing palettes. Free plan: a small foot credit is injected — suppressed by any authored `om-text` crediting OnlyMap, or by a paid key on a live frame's map. Deep reference: docs/cartograph.md.
 - Animation: `transition="get-fill-color 800ms, get-radius 400ms"` on a layer GPU-animates prop changes (also smooths streaming updates via `get-position`). Camera: the `fly-to` action takes `center`/`zoom`/`pitch`/`bearing`/`duration` (e.g. `duration="2s"`) — use it in behaviors or `data-emit` buttons; `zoom-to-feature` also accepts `duration`.
 - Pull-model frame rendering (external video frameworks such as Remotion): mark the map `data-om-recording` (the determinism switch — instant camera, no transitions/gesture interrupts, byte-stable `snapshot()`), then per frame `story.seek(t, {interpolateCamera: true})` + `await mapEl.whenSettled()` + capture; the same frame is byte-identical across processes and orderings. Story effect verbs (fade/pulse/trace/populate) evaluate as pure functions of story time under the switch — a seek landing mid-trace renders the half-drawn outline; only effects dispatched OUTSIDE a story (behaviors, ctx.emit) snap to their end state. `onlymapjs record` remains the built-in push-model path; `@nika-js/onlymap-remotion` packages the pull-model path for Remotion.
 - Travel-map animation: a `trace` story step takes `follow` (camera rides the drawing tip along the path) and `easing="linear|ease-in|ease-out|ease-in-out"`; combine with a TripsLayer route + pins toggled by later steps, and export with `npx onlymapjs record`.

package/onlymapjs.attributes.json

@@ -1196,6 +1196,7 @@
     "ScenegraphLayer": [
       "auto-highlight",
       "clip",
+      "color",
       "color-format",
       "coordinate-origin",
       "coordinate-system",