@woosh/meep-engine 3.60.4 → 3.60.6

package/src/shade/renderer/volumetrics/VOLUMETRIC_NOISE_PLAN_2026_10_15.md

@@ -1,336 +0,0 @@
-# Noise for participating-media volumes — design brief, 2026-10-15
-
-Status: **revision (c), built.** Written against the tree at 59861de99; the implementation is what
-the tree beside this file holds, and `NOTES.md` (2026-10-15) has the device numbers. Revisions (a)
-and (b) proposed a procedural field and then a texture with a stack of phases; Alex's answers
-(§6) settled on one static texture per slot in a general 3-D tile atlas, evolution by scrolling and
-a domain warp, no stacks, no clock. §4 and §5 describe what was built; §7 is what was left open.
-
-The ask: noise-modulated density on `ParticipatingMedia` volumes with a simple API and a cost that
-stays flat. The starting concept was a fixed number of octaves of 4-D noise, sampled in world
-space, with a per-volume scale and temporal rate, a per-volume lerp between uniform and noisy so one
-shader path serves both, and a global wind vector.
-
-## 1. What the tree had
-
-Facts, read from the sources, each of which shaped the design.
-
-**The volume model is a unit box posed by a matrix, carrying `density`, a particle spec, and an
-edge fade.** `PARTICIPATING_MEDIA_VOLUME_STRUCT` was `particle_spec`, `transform`,
-`transform_inverse`, `fade_distance`, `density` — a row in the scene database, rows packed from 0
-and rewritten whole whenever `SceneVolumetrics.version` moves. The renderer-side
-`ParticipatingMediaVolume` carried six dead stub fields (`noise_scale/amplitude/offset/velocity/
-evolution_speed`, `time`) from an older version of this very concept; nothing read them. Deleted.
-
-**The froxel density pass had no sample point.** `shader_volumetrics_build_participating_media`
-runs a 16×16 workgroup per z-slice over `ceil(w/16) × ceil(h/16) × 64` froxels — 120×68×64 =
-522,240 at 1080p, 2,073,600 at 2160p. Per froxel it builds the slab's world frustum (the z edges
-carry the per-column jitter), then for every volume row: reject on `aabb3_intersects_frustum`,
-estimate `frustum_unit_box_coverage` (8 three-plane intersections), evaluate the edge fade at the
-midpoint of the centre ray's intersection with the box, and accumulate `density × coverage × fade`.
-The density was a slab average, never a value at a point. Adding noise meant choosing a point, and
-the rule from 2026-10-09 applies — a froxel's value is evaluated at a per-frame jittered point,
-never a fixed lattice, or a large particle reading the volume shows the lattice as seams.
-
-**The lighting pass reads the density textures at a jittered point and the TAA integrates it.**
-STBN at the `(37, 71)` decorrelation offset, the 3-D TAA at 0.95, about 39 effective frames,
-reprojected by camera motion only. A density field that moves in world space is averaged over the
-window, not reprojected.
-
-**Nothing downstream re-evaluates the volumes.** The sun's optical-depth cascades and the
-local-light march both read `tParticipatingMedia1`, the froxel extinction texture. Noise costs once,
-in the density pass, never per shadow step.
-
-**The path tracer evaluates the volumes analytically and needs a majorant.**
-`participating_media_volume_density_at(volume, p)` is bounded above by `density`, and
-`chunk_medium_volumes`'s piece majorant is `Σ density × extinction`. The noise had to be a factor
-in [0, 1], applied inside that one function, or the null-collision estimator goes biased.
-
-**psrdnoise was in the tree with a GPU consumer already.** `chunk_psrdnoise3(x, period, alpha)`
-tiles at any integer period up to 289 on each axis. That property is what makes the bake exact:
-each octave gets an integer period across the tile, so each tiles on its own.
-
-**There was a 3-D slot atlas, a slot allocator, and a texture that resizes with its contents.**
-`GPUGeometrySDFManager` packs 50³ SDF slots into one 3-D texture and grows; `PageSlotAllocator` is
-an O(1) free list; `GPUTextureContext.resize(…, true)` copies the old extent across. None was a
-general tile atlas. The 2-D terrain atlas is a rect packer for arbitrary patches, a different thing.
-Brixelizer carries its own 8³ brick SDF atlas (the FFX port) and was left alone, as asked.
-
-**WebGPU has no bindless and no compressed 3-D textures.** A per-volume texture is an atlas or
-nothing, and the atlas is `r8unorm`: BC4 is 2-D only there. And no `r8unorm` storage texture in
-core WebGPU, which is why the bake writes a byte block into a storage buffer and copies it in.
-
-**Froxels are long in z.** On the 64-slice log curve a slab is ≈16% of its depth (≈8 m at 50 m,
-≈79 m at 500 m); a 16-px tile at 1080p is ≈1.4% of depth across. Features finer than the slab are
-averaged by the jitter and the TAA into a lower-contrast mean, so distant fog goes uniform on its own.
-
-## 2. What the field ships
-
-Tagged **VERIFIED** where the slides or source were read, **SECONDARY** where only documentation
-or a search summary was, **NOT RE-READ** where the material exists but was not opened.
-
-**Smolder — IO Interactive, 007 First Light, SIGGRAPH 2026 Advances (VERIFIED, slides + notes).**
-Not a noise system: a per-pixel raymarcher over baked animated density (OpenVDB frames in a custom
-sparse brick format, 4³ bricks stored 5³, 4 bits a voxel), 4-moment volumetric shadows on async
-compute, segment-based overlap culling (4 volumes), an adaptive lighting rate. No temporal
-filtering inside, blue noise per pixel, FSR's reactive mask. The density lookup "was only a small
-fraction" of the per-sample cost; and the one thing they refused was a fallback that makes artists
-"author content twice". 450 volumes shipped. Slides:
-`advances.realtimerendering.com/s2026/content/AlexanderMueller_Smolder_Siggraph26.pdf`.
-
-**Enshrouded — Keen Games, GPC 2025 "The fog is lifting" (VERIFIED, slides).** A froxel volume
-(160×90×64, to 10 km, temporally accumulated on the populate pass only) for atmosphere, weather and
-VFX-injected box volumes — `densityTexture, uvOffset, uvScale, density, falloff, …` — and a
-half-resolution ray march for the dense shroud over a world voxel volume (16³ tiles with a
-"half voxel border on each side for seamless trilinear filtering"). **Detail noise is baked**:
-Nubis-Cubed "Curly-Alligator" and "Alligator" types, "tileable, 128x128x128 four channels
-uncompressed", eroding density at ray-march samples. Slides on
-`static.graphicsprogrammingconference.com/public/2025/talks/the-fog-is-lifting-volumetric-rendering-enshrouded/`.
-The GPC 2024 talk (Lukas Feller, linked at the top of `NOTES.md`) is video only.
-
-**Assassin's Creed 4 — Wronski, SIGGRAPH 2014 (VERIFIED).** "Just one octave of Perlin noise that
-is animated by wind. We tried using multiple octaves, but in the end difference was rather subtle
-for added cost." Density and lighting 0.43 ms on PS4 at 160×90×64.
-
-**Bevy `scrolling_fog` (VERIFIED, source).** A repeating 3-D noise texture, a
-`density_texture_offset` the app advances on the CPU by `velocity × dt`, jitter under TAA.
-
-**Godot 4 FogMaterial (SECONDARY).** A static 3-D `density_texture` in clamped local box space,
-height falloff, edge fade; animation needs a custom shader. **Unity HDRP (SECONDARY):** 3-D mask
-texture, scroll speed and tiling per axis. **Unreal 5.4 (VERIFIED):** local fog
-volumes are analytic (radial + exponential height, no noise); froxel fog density comes from
-Volume-domain materials evaluated per froxel at a jittered slice depth. **Frostbite 2015, RDR2
-2019, Nubis Cubed 2023 (NOT RE-READ):** PPTX only, or cited through Enshrouded.
-
-**The pattern.** Froxel fog ships one or two octaves of 3-D noise scrolled by wind, or a baked
-tileable 3-D texture eroding a density; hero effects are baked animated voxels. The 3-D texture is
-the universal artist hook, and 128³ is the size the two most recent references use.
-
-## 3. What it costs
-
-Per (froxel, volume) the noise is two trilinear `r8unorm` taps into an atlas of a few MiB (one tap
-when `noise_warp == 0`) plus ~25 scalar ops of coordinate arithmetic, against the ~300 ops the
-coverage estimator already spends per volume in the same loop. Paid only inside `aabb_intersects`;
-a froxel with no volume pays nothing, and a volume at `noise_strength == 0` branches around it
-uniformly, so a scene with no noisy volume costs what it did before. The device numbers, taken by
-the paired protocol in `volumetrics_froxel/measure.js` at 2160p, are in `NOTES.md` (2026-10-15).
-
-## 4. The concept, item by item, as built
-
-**Fixed number of octaves — 3, baked into the default tile.** Lacunarity 2, gain 0.5, three
-octaves of `psrdnoise3` at integer periods 4, 8 and 16 lattice units across the content, each
-offset by a constant times its index so the doubled lattices do not sit on the base one's corners.
-The finest octave is 8 texels a wavelength in a 128³ tile. At runtime the octave count is not a
-parameter: it is whatever the texture holds, and a user tile may hold one or ten at the same cost.
-
-**4-D noise — dropped (F1), on Alex's call.** "3-D preserves mass/volume better perceptually; 4-D
-is great for variation but with discrete samples not worth it, blurrier borders for little benefit
-and 2x taps." No phase stack either, for memory. The evolution is **scroll plus a domain warp**: a
-second tap of the same tile at half frequency and its own, slower scroll displaces the base lookup
-along a fixed oblique direction by `noise_warp` tile units times how far the warp value sits from
-its mean. As the two scrolls slide against each other (ratio 0.37) the base field changes shape in
-place. `noise_warp == 0` is one tap and a field that only slides.
-
-**Time tick / supplied time — no clock at all.** The scroll is a **phase on the volume**, two
-`[0, 1)³` offsets (`noise_offset`, `noise_warp_offset`) advanced once a frame by
-`participating_media_volume_advance_noise(volume, velocity, dt)` in f64 and wrapped there — a
-running sum, so a change of wind never jumps the field and a session running for days never grows
-a number. `ParticipatingMediaSystem.update(dt)` is the caller for an ECS scene; a playground calls
-it from its loop. Not called, the field stands still, which is what an accumulating reference
-wants. The GPU never multiplies a velocity by a time.
-
-**Temporal scale per volume — became `noise_warp`.** A rate had no meaning without a 4th
-dimension; the warp amount is the knob that sets how much the field changes shape as it moves.
-
-**World-space sampling — kept, with repeat.** The tile repeats every `noise_scale` metres along
-each world axis; `fract` folds the world position onto the tile. Neighbouring volumes with the same
-settings share one continuous field, and a volume carried by a moving entity has the world's fog
-flow through it. User tiles must tile seamlessly; the engine's does by construction.
-
-**Noise scale per volume — one scalar, metres per tile repeat.** `noise_scale`, default 32 m,
-which with the default tile's four base features puts them 8 m apart and the finest octave at 2 m.
-
-**Lerp between noisy and uniform — `noise_strength ∈ [0, 1]`, blending to the texture's [0, 1]
-value.** `density(p) = density × fade(p) × mix(1, tile(p), strength)`. Bounded by `density`, so the
-tracer's majorant, the ray spans and the froxel coverage arithmetic are untouched. The remap that
-shapes the wisps — `smoothstep(−0.6, 0.6, fBm)`, psrdnoise's fBm sits in about ±0.7 — is baked, so
-the texture *is* the multiplier and a user tile is one too, with nothing more to say.
-
-**Uniform path.** `noise_strength == 0` branches around the taps inside the volume loop: a property
-of the volume, identical for every thread, so a uniform branch and not a variant.
-
-**Global wind — `SceneVolumetrics.wind`, felt per medium.** Per Alex's F7: effective velocity
-`= wind × noise_wind_factor + noise_velocity`, two explicit fields on the component, folded by the
-system into the phase. The row carries phases only. *Superseded 2026-10-16 (GPU fluid S3,
-`GPU_FLUID_CASCADES_PROPOSAL_2026_10_16.md` §6): `SceneVolumetrics.wind` is removed, and a scroll
-writes each row's phase words through `SceneVolumetrics.invalidate_noise()` instead of repacking the
-table. The phases fold the scene's far field (`GPUFluidSystem.far_field`, still air without one);
-the fluid's departure from it is sampled on the GPU wherever the noise is evaluated and carries the
-noise as a flow map, on a flow clock the host folds into the row (`noise_flow_phase`, period 2.5 s),
-scaled by `noise_wind_factor`, which the row now carries too. A per-volume read-back of the wind was
-built first and withdrawn on Alex's word: the data is on the GPU. The scroll also changed sign: it
-had carried the noise upwind.*
-
-**Where the froxel pass evaluates it.** Once per froxel, before the volume loop, a world point
-inside the slab at an STBN jitter with a **third** decorrelation offset `(53, 97)` — the aerial pass
-owns zero and the lighting pass `(37, 71)`, and two passes drawing the same number lock their
-nuisance variables together, which biases the estimate and no TAA window recovers it. Per volume
-inside `aabb_intersects`, the multiplier at that point scales the slab's weight. The slab's z jitter
-is already per frame, so the point moves on all three axes.
-
-**Band limiting — none (F2).** No mips (a one-texel skirt is wrong from level 1 down, and the
-tracer samples one level anyway), no octave skipping; the texture is band-limited at its texel and
-the jitter plus the TAA average the froxel.
-
-**What the path tracer gets.** The same `participating_media_volume_density_at`, which now needs
-`tParticipatingMediaNoise` and `sParticipatingMediaNoise` bound by every shader that imports it:
-the accumulating tracer, `graph_path_trace_scene`, the Brick4 probe bake, both lightmap viz bakes
-and the deprecated LPV trace all declare and bind them.
-
-**The TAA envelope — a stated limit.** At 0.95 the history spans ≈39 frames, 0.65 s at 60 Hz; a
-feature scrolls `|v| × 0.65 s` inside that window. Wind at a few m/s on 8 m features is fine; 10 m/s
-on 2 m features is mush, and a history reprojected by the scroll is a separate piece of work.
-
-## 5. The design, in one place
-
-### 5.1 `GPUTileAtlas3D` — general, one instance for the volumetrics
-
-`shade/renderer/texture/atlas/GPUTileAtlas3D.js`. One 3-D `GPUTextureContext` of one format, cut
-into tiles of `content_size + 2·border` texels, slots linear (x fastest, z slowest), handed out by
-a free list that reuses the most recently freed slot first.
-
-- `new GPUTileAtlas3D({ device, format, content_size, border, slots_x, slots_y, capacity, usage, label })`,
-  every argument required. `slots_x`/`slots_y` are the starting footprint; `capacity` is rounded
-  up to whole z layers of it. `usage` is how the clients bind the texture (`TEXTURE_BINDING`,
-  `STORAGE_BINDING`); the atlas adds the `COPY_DST | COPY_SRC` its writes and grows need (added
-  2026-10-16 for the SDF manager, whose build stores into its tiles).
-- `allocate()` returns a slot and, when none is free, **grows the texture by a layer**, copying the
-  old extent across (`resize(…, true)`; x and y do not change, so every tile keeps its origin).
-  **Once z is at `maxTextureDimension3D` it widens instead** (2026-10-16, Alex's design): the
-  smaller of x and y doubles (x on a tie, capped at the device), the layers halve (+1 for the slot
-  being allocated), and every tile handed out is copied — in runs along x, a whole old row per
-  copy when a side doubles — to where its slot sits under the new footprint, through
-  `GPUTextureContext.replace`. A slot keeps its index and texels; its origin moves. Rare and as
-  costly as a full copy. The ceiling is `capacity_limit`, the device's limit in tiles cubed; past
-  it, `allocate` throws with the numbers. `free(slot)`, `is_live(slot)`, `live_count`.
-- `write(slot, content)` takes `content_size³` texels and builds the skirt on the host
-  (`tile_atlas_bordered_block`, pure and exported: the skirt is the content wrapped) in one
-  `writeTexture`. `copy_from_buffer(encoder, slot, buffer, offset, bytes_per_row)` lands a whole
-  bordered tile a device producer wrote, rows padded to the copy alignment.
-- `slot_origin(slot, out)` is the same arithmetic in JS and WGSL.
-- A grow replaces the `GPUTexture` and a widen moves every tile; take a view, the footprint or an
-  origin after allocating, never across an `allocate`. The frame graph imports the texture per
-  frame and the noise shader takes the footprint from `textureDimensions`, so the passes are fine
-  by construction.
-
-The volumetrics instance (`PARTICIPATING_MEDIA_NOISE.js`): `r8unorm`, content 128, border 1
-(tile 130), footprint 2×1, capacity 2 — the engine's tile and one spare, 4.4 MB — growing by 2
-slots (4.4 MB) a layer to 30 on a 2048-limit device, then widening. Per scene, on `GPUSceneVolumetrics.noise_atlas`:
-a volume names its tile by slot and the volumes are the scene's.
-
-`GPUGeometrySDFManager` is the second instance (G4, 2026-10-16): `r32float`, content 50, border 0
-(every reader loads inside its tile and clamps the far tap to the last texel, and a wrapped skirt
-would be wrong for a distance field anyway), starting footprint 4×4 — the 16 slots it started with,
-and its smallest growth step — to 40 layers (640 slots, 320 MB) on a 2048-limit device, then
-widening to 8×4×21 and on. A slot is its geometry's id, so it claims slots in order and frees none.
-Brixelizer's atlas is not migrated.
-
-### 5.2 The engine's default tile — slot 0, baked on first update
-
-`shader_volumetrics_noise_tile_bake` + `bake_volumetrics_noise_tile`: a compute dispatch over the
-whole 130³ tile — skirt included, by wrapping the texel index, the same rule the host applies — into
-a transient storage buffer of 256-byte rows, four `r8unorm` texels a word, then
-`copyBufferToTexture` into the slot, in one command stream. Three octaves at periods 4/8/16,
-`features` 4, remap baked. Runs once, from `GPUSceneVolumetrics.update` the first time it is
-called (a bake needs a command stream and the constructor has none). Deterministic, so the
-emulator bakes small tiles in `shader_volumetrics_noise_tile_bake.spec.js`.
-
-### 5.3 The API
-
-ECS, flat on `ParticipatingMedia`:
-
-    noise_strength    = 0        // [0, 1]; 0 = uniform, nothing evaluated
-    noise_scale       = 32       // metres per tile repeat (default tile: 8 m base features)
-    noise_warp        = 0.1      // tile units the warp may displace by; 0 = one tap, slide only
-    noise_velocity    = [0,0,0]  // m/s, the medium's own motion, added to the wind
-    noise_wind_factor = 1        // how much of the scene wind it feels
-    noise_texture     = 0        // atlas slot; 0 = the engine's
-
-Scene: `SceneVolumetrics.wind` (m/s) — removed 2026-10-16, see above. Tiles: `GPUSceneVolumetrics.noise_atlas` —
-`allocate()`, `write(slot, Uint8Array(128³))`, `free(slot)`.
-
-Renderer-side `ParticipatingMediaVolume`: `noise_strength`, `noise_scale`, `noise_warp`,
-`noise_slot`, and the phase state `noise_offset`, `noise_warp_offset`, advanced by
-`participating_media_volume_advance_noise(volume, velocity_m_s, dt_s)`, which returns whether
-anything moved.
-
-GPU row, appended to `PARTICIPATING_MEDIA_VOLUME_STRUCT`: `noise_offset: vec3f`,
-`noise_warp_offset: vec3f`, `noise_slot: u32`, `noise_strength: f32`, `noise_inverse_scale: f32`,
-`noise_warp: f32`. Packed by `GPUSceneVolumetrics` whenever the version moves — which, while any
-volume's phase moves, is every frame (tens of rows). *Since 2026-10-16 a scroll moves
-`SceneVolumetrics.noise_version` instead, and only the moved rows' phase words go up: seven, with
-`noise_flow_phase`; `noise_wind_factor` and `noise_flow_phase` sit in the padding after `density`.*
-
-Serialization: `ParticipatingMediaSerializationAdapter` version 1 (eighteen doubles),
-`ParticipatingMediaSerializationUpgrader_0_1` fills a version-0 save with the defaults.
-
-### 5.4 The shader
-
-`chunk_participating_media_noise`, bound as `tParticipatingMediaNoise` (`texture_3d<f32>`) with
-`sParticipatingMediaNoise` (linear, clamped — the wrap is the skirt's and `fract`'s, never the
-sampler's, which would wrap the whole atlas):
-
-    tap(slot, u):  cell = (slot mod sx, (slot / sx) mod sy, slot / (sx·sy)), sx/sy from textureDimensions
-                   texel = cell·130 + 1 + fract(u)·128;  sample at texel / atlas_size
-    multiplier(volume, p):
-                   base = p · noise_inverse_scale;  u = base + noise_offset
-                   if noise_warp > 0:  w = tap(slot, base·0.5 + noise_warp_offset)
-                                       u += (w − 0.5) · noise_warp · (0.6, 0.48, −0.64)
-                   return mix(1, tap(slot, u), noise_strength)
-
-`participating_media_volume_density_at` multiplies it in under `if (noise_strength > 0)`.
-
-### 5.5 Specs
-
-`GPUTileAtlas3D.spec.js` (slots, placement, skirt, growth with tiles preserved — read back off the
-software device — the limit, formats); `shader_volumetrics_noise_tile_bake.spec.js` (emulator: skirt
-= content wrapped, range, periodicity of the field across the content, padding words untouched,
-octaves and features, determinism, the octave ceiling, WGSL validity);
-`chunk_participating_media_noise.spec.js` (emulator, hand-built atlas: texel centres, wrap, the
-skirt blend and never the next slot, strength, scale, scroll, the warp's displacement in closed
-form, the density function bounded); `ParticipatingMediaVolume.spec.js` (the advance: units,
-wrap, continuity, a week of frames); `GPUSceneVolumetrics.spec.js` and
-`GPUViewVolumetrics.passes.spec.js` (unchanged, now encoding the atlas); the ECS `*.noise.spec.js`
-trio (component JSON/equality, adapter round trip and the 0→1 upgrade, the system's fold and
-per-frame invalidation).
-
-## 6. The forks, as answered
-
-Revision (a): **F1** flow noise — yes in principle, then ditched with the texture pivot (§4).
-**F2** octave band-limiting — "with textures insignificant", dropped. **F3** [0, 1] multiplier —
-"absolutely". **F4** CPU-side wrapped phase — fine, "a mod on the CPU". **F5** scalar scale — fine.
-**F6** delete the stubs — yes. **F7** per-volume velocity additive plus a multiplier on the wind —
-yes: `noise_velocity`, `noise_wind_factor`.
-
-Revision (b): **G1** no stacks — "one static texture, temporal variation via domain warping and
-scrolling"; built as scroll + warp. **G2** 128³ — yes. **G3** general atlas from day one — yes.
-**G4** migrating the SDF manager — a follow-up; Brixelizer left alone or promoted out of its FFX
-naming, not built on. **G5** world-space repeat only — yes. **G6** automatic growth — "house style";
-16 slots was fine, 2 was chosen as the first layer since a layer is 4.4 MB and growth is automatic.
-**G7** `r8unorm` — yes.
-
-## 7. Open, by measurement or by later ask
-
-- The repeat distance is `noise_scale`; at 32 m on a wide bank seen from above it may read as a
-  lattice. The fix, if it shows, is a user tile with more features or a larger scale — no code.
-- A per-volume `contrast`, a local clamped sampling mode (a texture that *is* the shape), exponential
-  height falloff, and history reprojected by the scroll: each a one-field or one-pass addition the
-  layout leaves room for, none asked for.
-- The froxel-versus-tracer parity check for a noisy volume was not run on the device this pass;
-  both paths read one function and the chunk spec pins the function.
-
-## 8. Not verified this pass
-
-RDR2's and Frostbite's fog-density slides (PPTX only). Nubis Cubed's noise authoring (cited
-through Enshrouded). The GPC 2024 Enshrouded talk (video only). The contrast dip of a phase
-crossfade, which the no-stacks decision made moot.

package/src/shade/renderer/water/medium/GPUWaterBodies.d.ts

@@ -1,23 +0,0 @@
-/** A view's water bodies: which the view draws and how, and their meshlets in row order. */
-export class GPUWaterBodies {
-    /**
-     * Draw culling over the frame's visible list (WATER_SURFACE_DISTANCE_LOD_PROPOSAL_2026_10_06.md §4.3), each drawn
-     * body's lattice frame at its row, and the drawn bodies' meshlets. The expansion is an ordered prefix scan over the
-     * rows, which preserves equal-depth ownership, including after holes and reuse.
-     *
-     * @returns {{meshlets: number, lod_bodies: number}} resources
-     */
-    meshlets({ graph, view, camera, meshlet_capacity, visible_rows, lod, steep }: {
-        graph: any;
-        view: any;
-        camera: any;
-        meshlet_capacity: any;
-        visible_rows: any;
-        lod: any;
-        steep: any;
-    }): {
-        meshlets: number;
-        lod_bodies: number;
-    };
-}
-//# sourceMappingURL=GPUWaterBodies.d.ts.map

package/src/shade/renderer/water/medium/GPUWaterBodies.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"GPUWaterBodies.d.ts","sourceRoot":"","sources":["../../../../../../src/shade/renderer/water/medium/GPUWaterBodies.js"],"names":[],"mappings":"AAkBA,4FAA4F;AAC5F;IACI;;;;;;OAMG;IACH;;;;;;;;QAFa;QAAC,QAAQ,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,MAAM,CAAA;KAAC,CAgDlD;CACJ"}

package/src/shade/renderer/water/medium/GPUWaterBodies.js

@@ -1,75 +0,0 @@
-import { BufferResourceDescriptor } from "../../../device/graph/BufferResourceDescriptor.js";
-import { round_up } from "../../../../core/math/round_up.js";
-import {
-    graph_prefix_scan_csdldf,
-    prefix_scan_csdldf_output_descriptor
-} from "../../gpu_primitive/prefix_sum/v1/graph_prefix_scan_csdldf.js";
-import { graph_create_buffer } from "../../shader/graph/graph_create_buffer.js";
-import {
-    build_indirect_compute_dispatch_command,
-    INDIRECT_COMPUTE_COMMAND_DESCRIPTOR
-} from "../../gpu_primitive/indirect/IndirectComputeCommandBuildPass.js";
-import { shader_prefix_sum_to_command } from "../../rasterize/expand/mesh/shader_prefix_sum_to_command.js";
-import { shader_increment_count_by_prefix_sum } from "../../rasterize/expand/mesh/shader_increment_count_by_prefix_sum.js";
-import { graph_compute_pass } from "../../shader/graph/compute/graph_compute_pass.js";
-import { shader_water_meshlets_expand } from "./shader_water_meshlets.js";
-import { shader_water_lod_bodies } from "../surface/lod/shader_water_lod.js";
-import { WATER_LOD_BODY_STRUCT, WATER_LOD_VIEW_STRUCT } from "../surface/lod/WATER_LOD.js";
-
-/** A view's water bodies: which the view draws and how, and their meshlets in row order. */
-export class GPUWaterBodies {
-    /**
-     * Draw culling over the frame's visible list (WATER_SURFACE_DISTANCE_LOD_PROPOSAL_2026_10_06.md §4.3), each drawn
-     * body's lattice frame at its row, and the drawn bodies' meshlets. The expansion is an ordered prefix scan over the
-     * rows, which preserves equal-depth ownership, including after holes and reuse.
-     *
-     * @returns {{meshlets: number, lod_bodies: number}} resources
-     */
-    meshlets({ graph, view, camera, meshlet_capacity, visible_rows, lod, steep }) {
-        const scene = view.scene, table = scene.scene_database.get('water_bodies');
-        const capacity = Math.max(table.element_capacity, 1);
-        const data = {};
-        const pass = graph.add('Water / lod bodies', data, (d, resources, { encoder }) => {
-            const command = resources.get(d.command);
-            build_indirect_compute_dispatch_command(64, encoder, visible_rows, 0, command, 0, true);
-            shader_water_lod_bodies.dispatchIndirect({ encoder, command, bindings: {
-                capacity: round_up(capacity, 4), lod: encoder.allocateTransientValueBuffer(WATER_LOD_VIEW_STRUCT, lod),
-                camera: resources.get(d.camera), scene_database: scene.scene_database_buffer,
-                geometries: scene.geometries.buffer_metadata, visible_rows, steep,
-                lod_bodies: resources.get(d.lod_bodies), counts: resources.get(d.counts),
-            } });
-        });
-        data.camera = pass.read(camera);
-        // The shared scan reads vec4u blocks, including the padded final block.
-        const bytes = 16 + round_up(capacity, 4) * 4;
-        data.counts = pass.create('counts', BufferResourceDescriptor.from(bytes, GPUBufferUsage.STORAGE, 0, bytes));
-        data.command = pass.create('Water lod bodies command', INDIRECT_COMPUTE_COMMAND_DESCRIPTOR);
-        data.lod_bodies = pass.create('Water lod bodies', BufferResourceDescriptor.from(capacity * WATER_LOD_BODY_STRUCT.size,
-            GPUBufferUsage.STORAGE));
-        const sums = graph_prefix_scan_csdldf({ graph, input: data.counts,
-            output: graph_create_buffer(graph, 'Water lod body meshlet sums', prefix_scan_csdldf_output_descriptor(capacity)),
-            element_max_count: capacity });
-        const command_data = {};
-        const command_pass = graph.add('Water / expand command', command_data, (d, resources, { encoder }) => {
-            shader_prefix_sum_to_command.dispatch({ encoder, bindings: {
-                prefix_sums: resources.get(d.sums), command: resources.get(d.command),
-            } });
-        });
-        command_data.sums = command_pass.read(sums);
-        const command = command_data.command = command_pass.create('command', INDIRECT_COMPUTE_COMMAND_DESCRIPTOR);
-        const expand_data = {};
-        const expand = graph.add('Water / expand meshlets', expand_data, (d, resources, { encoder }) => {
-            shader_water_meshlets_expand.dispatchIndirect({ encoder, command: resources.get(d.command), bindings: {
-                prefix_sums: resources.get(d.sums), output: resources.get(d.output),
-                scene_database: scene.scene_database_buffer, geometries: scene.geometries.buffer_metadata,
-            } });
-        });
-        expand_data.command = expand.read(command);
-        expand_data.sums = expand.read(sums);
-        expand_data.output = expand.create('meshlets', BufferResourceDescriptor.from(16 + Math.max(meshlet_capacity, 1) * 8,
-            GPUBufferUsage.STORAGE, 0, 16));
-        const { output } = graph_compute_pass({ graph, shader: shader_increment_count_by_prefix_sum,
-            bindings: { prefix_sums: sums, output: expand_data.output } });
-        return { meshlets: output, lod_bodies: data.lod_bodies };
-    }
-}

package/src/shade/wgsl/NOTES.md

@@ -1,4 +0,0 @@
-WASM is taken from this
-commit : https://github.com/gpuweb/tree-sitter-wgsl/blob/ca727304343930195bcaf36b97d23431d10fc996/tree-sitter-wgsl.wasm
-
-THE official grammer is tree-sitter, from WebGPU W3C: https://github.com/gpuweb/tree-sitter-wgsl