The new opt-in path is flux.schema = "mjolnir-flux-v2" with domains.
Existing single-domain/coast maps retain their previous loader behavior.
Authoring uses playback: "simulate"; release assembly requires
playback: "baked" and a matching saved state. Baked playback spawns a native
surface (River by default), imports the state textures, and calls the installed
SetSimulationState(State) function. No simulation domain or source modifiers
are created by that path.
Cooked Blueprint inspection matters here: SetSimulationState assigns the state
and invokes InitializeStates. This game's StopSimulation talks only to its
editor helper, so calling it is not evidence of stopped runtime simulation.
The author documents direct surface playback from a saved state, and recommends
bounded simulation resolution and capture-only slopes for difficult cascades:
Fluid Flux documentation.
Inputs and capture geometry
Upgrade config flux.heightfield names the decoded runtime master, with
surface_interpolation: "triangle_nw_se"; flux.water_geometry names the 3D
water/profile geometry. Each entry in flux.domains contains id, domain,
sources, and optionally playback, prime_ms, surface, color_preset, and
surface_vectors, surface_scalars, and native simulation foam parameters.
Domain positions, cell size and height use Unreal centimetres;
the master uses authoring metres with Y reflected into Unreal.
tools/level/flux_heightfield.py retains original grid vertices and the same
diagonal as Landscape, without resampling. It writes a double-sided,
capture-only mesh plus a receipt. Domain rectangles outside actual terrain
coverage fail rather than inventing a floor. The helper rejects resolutions
above 1024 or non-power-of-two dimensions and capture height above 32,000 cm.
This is a capture mesh, not an additional Blam collision mesh.
The current fork's Flux capture does not reliably capture Landscape. The older
CE-based flux_bed path therefore cannot be reused for PWR's buried coarse core.
The new bed is built from actual rendered terrain. Bridges, foliage and other
overhead decor are excluded from capture; the capture bed is hidden afterward.
The new path sets the native SceneCapture to UseShowOnlyList and populates
ShowOnlyActorComponents from the explicit bed IDs. A tag exclusion list is
insufficient: streamed terrain can exist outside the loader's actor dictionary.
This was caught by comparing every captured pixel with the final field while
authoring v5 water inside the older v3 visual world. Before the allowlist, old
mountains entered capture; afterward every pixel matched within the exported
half-float depth precision. Do not replace this with a broader scene capture.
Bake validity and playback
Five SHA-256 values bind a bake to its terrain manifest, decoded terrain data,
3D water geometry, capture mesh, and domain/source/simulation settings. The settings
hash also binds the explicit-bed capture policy. Domain geometry and
canvas origin must match as well. Assembly validates EXR headers/dimensions and
seals the texture hashes. A new terrain with unchanged domain bounds still
invalidates the old bake. All named states live under flux_state/<id>/ and
contain state.json, height.exr, velocity.exr, and ground.exr.
Ground is required by the native surface material, even without a solver.
UE5.5's Kismet exporter writes EXR only for RGBA16F targets, so the height and
ground maps are copied through RGBA16F before export. The source capture mesh
retains exact terrain vertices; the exported depth has half-float quantization.
All-baked assembly omits the capture bed actor entirely. Its mesh and receipt
remain offline inputs for provenance validation; they are not runtime surfaces.
Authoring commands are mjolnir_flux_status and
mjolnir_flux_bake <id|all>. Baking records current GPU state; removal of primers
does not certify settled or visually correct flow. Review shoreline depth,
cascade continuity and bridge clearance first, then export, assemble with
playback: "baked", reload and confirm simulation_domain:false.
Release packaging rejects authoring modes or any missing named state.
Fluid Flux remains presentation here. Native player and vehicle support stays with the Blam bed, ford and bridge collision. This change does not implement swimming, drowning, buoyancy, or replicated fluid forces.
Temporary proof and tests
levels/warfare/pwr/terrain/prepare_flux_probe.py --heightfield <runtime-master> --geometry <terrain.geometry.json> --out <proof-directory> --package builds an
isolated 512 m upper-cascade capture tile. It does not install or call the game.
tools/level/flux_probe_install.py install|restore --proof <proof-directory>
requires a closed game, snapshots exact installed metadata/helper hashes, and
touches only narrow loader hooks plus temporary chunk 962. No DLL or main map
container is replaced. tools/level/flux_probe_readback.lua reads actual surface,
solver and capture-bed transforms through the existing bridge.
Offline checks:
lua tools/tests/test_flux_runtime.lua
python -m unittest discover -s tools/level -p test_flux_heightfield.py
python -m unittest discover -s tools/level -p test_upgrade_ce_map.py
python -m unittest discover -s tools/level -p test_pack_local_map.py
Tests cover exact source vertices/diagonal, stale terrain/water/settings, modified capture/EXR files, missing bakes, two independent baked surfaces with zero solver/modifier allocation, invalid paths, failure cleanup and stale callbacks and the explicit capture-component allowlist.
The temporary v3 upper-cascade proof completed three-texture saved-state reload:
zero domains, zero source modifiers, one native river surface. Its source-bed
vertex error was zero; exported capture depth differed by at most 0.1625 m.
out/pwr/flux-proof-v3/final-baked-cascade.png and the later screenshot show
the same saved surface. The original installed helper/metadata and camera poses
were restored by hash, and the temporary container removed with the game closed.
That proof is not relabelled as a v5 state.
PWR final-source domains
levels/warfare/pwr/terrain/plan_flux.py derives capture planes from the accepted
runtime grid and writes the isolated capture package, authoring spec, release
config fragment, and receipt. It does not install or call the game.
python levels/warfare/pwr/terrain/plan_flux.py `
--heightfield levels/warfare/pwr/terrain/runtime-v5/runtime-master.json `
--geometry levels/warfare/pwr/terrain/final-v5/terrain.geometry.json `
--out out/pwr/flux-v5 --package
The accepted manifest is cfa587ac8c9bc565270c035ee3a281b8611c8197fb7c6c0833990b1b6c4ee45f.
Two 1024² domains use 0.9765625 m cells. The upper rectangle covers authored
X[-732,268], Y[0,1000] m; the lower covers X[-232,768], Y[-1000,0] m.
The upper capture plane/range is 295/282 m; the lower is 130/116 m. Both keep
the capture reference below the minimum bed and the camera above the maximum.
The grids align at Y=0, with a shared authored head of 92.22669 m. Native Adjust
controls hold that head and the 55 m lake during authoring. They are removed
with the solver for saved-state playback. This does not imply automatic fluid
transfer between domains.
Lake controls stay 90 m from elevated river mouths: a broad modifier holding 55 m across a higher channel drains it and creates a dry gap. Primers are removed after 60 seconds; steady inlet/seam/lake controls remain only for baking. Native foam generation is restrained so calm reaches stay teal and cascades retain localized highlights. Free flow alone settled roughly 1.5–1.7 m below the authored profile in the initial proof; authored targets are not measurements. Narrow inlet controls additionally require their entire footprint plus margin to stay below the lake level. They fill the reservoir closer to incoming falls without draining the elevated channel above them.
tools/level/audit_flux_state.py reads the exported textures and reports all-pixel
capture error, centreline wetness, lake level and seam differences. It requires
NumPy and OpenEXR (tools/level/requirements-flux.txt).
The final v5-source bake and a fresh-process playback proof completed on
2026-11-09. Evidence is under out/pwr/flux-v5/:
final/flux_state/contains the two sealed states, with six EXRs.final/receipt.jsonhashes every file;final/flux.spec.jsonembeds matching provenance. The states were copied to the final upgrade work folder.final.audit.jsonmeasures all 2,097,152 captured pixels. Maximum absolute error is 0.16292 m in the upper tile and 0.08 m in the lower tile, from the half-float exported depth. The capture mesh retains the exact runtime grid (floating-point comparison error about 1e-12 m).- All 1,245 main-river/braid centreline samples are wet. The adjacent seam edges differ by 0.035 m, with no dry seam samples. The calm reservoir's 201,820 measured pixels are within -0.07/+0.01 m of 55 m. Turbulent inlet pockets are reported separately: the whole lake footprint ranges -2.055/+0.22 m from 55 m. The simulated flowing reaches also settle below the design profiles; the lower final cascade has a local maximum profile difference of 6.61 m. These are fluid results, not terrain/collision errors or a claim of flat water everywhere.
clean-start.readback.txtandclean-start-later.readback.txtobserve the PWR world in a new game process: two native river surfaces, zero simulation domains, zero modifier sources, zero capture-bed actors. All six input maps areTexture2D, not render targets. The temporary capture container was physically removed before that launch.clean-start-cascade.png,clean-start-lake.png,clean-start-mouth.png, andclean-start-seam.pngshow the saved-state water inside the older installed visual world.cleanup.jsonverifies the original installed helper/metadata hashes and camera poses were restored, generated proof states removed, game closed, and temporary container absent. No DLL or main map container was replaced.
tools/level/collect_flux_bake.py seals completed native exports into a fresh
output directory and produces the baked spec without changing installed files.
The final assembled v5 world's appearance, remote behavior and performance
remain separate acceptance gates. Fluid Flux forces still do not drive Blam
players or vehicles.
Optional baked lake waves
A baked domain can opt into the shipped Coastline material and a native
BP_FluxOceanWave actor. This is presentation over the existing saved state:
it does not change its bounds, source controls, provenance, or six EXR files.
Merge out/pwr/flux-v5/lower-lake-wave.fragment.json into the existing
lower_lake entry; it is a partial fragment, not a replacement domain.
{
"surface": "/Game/FluidFlux/Surface/Templates/Coastline/BP_FluxSurface_Coastline.BP_FluxSurface_Coastline_C",
"ocean_wave": {
"height_cm": 18,
"length_cm": 1800,
"choppiness": 0,
"offset_cm": 0,
"water_height_cm": 5500
}
}
The hook requires baked playback and that exact Coastline class. All five numbers are mandatory and finite; unknown keys fail before any actor or delayed callback is created. Height is bounded to 0.01–50 cm, wavelength to 100–10,000 cm, choppiness to 0–1, and vertical offset to ±50 cm. The water plane plus offset must stay inside the domain's captured vertical range. Its XY follows the domain centre; the canvas origin is added once to XYZ. World position must stay inside the native Flux coordinate radius.
Before native construction finishes, the wave receives UseRenderTarget=false
and the explicit parameters. Coastline receives SurfaceMeshMode=5, the baked
state and OceanWaveActor before its construction finishes. The installed
volume texture animates in the native material; the wave actor does not tick or
create its own render target. Coastline's additional detail displacement is
disabled through WaveDetailWPOScale=0, including after delayed colour setup.
Normal-map ripples remain. These are native amplitude controls, not an
analytical proof that every shaded vertex stays within ±18 cm.
Both actors enter the loader's ordinary actor registry. Partial startup failure destroys them, and reload/travel invalidates pending work. No per-frame Lua, solver domain, source modifier, capture bed, or extra fluid bake is introduced. Offline tests exercise construction ordering, transformed coordinates, failure cleanup, the actual loader reload cleanup, invalid option rejection, and unchanged simulation provenance. The default River path remains covered.
The 2026-11-09 temporary native probe showed the finite lake and islands with
the existing lower waterfall still visible. Evidence is in
out/pwr/visual-v5-qa/pwr_v5_water_wave_close-2026-11-09T17-25-24-161Z.png and
pwr_v5_wave_footprint-2026-11-09T17-26-50-388Z.png. Native readback confirmed
UseRenderTarget=false, VT_OceanWave, and disabled actor tick. The temporary
actors were removed and the original surface restored afterward. A final fresh
process using the production hook still needs its own receipt.
Execute tools/level/flux_wave_readback.lua through the existing read-only game
bridge after that launch. It requires zero domains/sources/beds, one River,
one Coastline, one non-ticking wave, finite mesh mode 5, three linear Texture2D
maps per state, the expected wave settings and installed VolumeTexture. It
prints surface/state/wave transforms for review. Whole-domain shore, island and
upper/lower seam screenshots remain separate visual acceptance checks.
The official surface template documentation describes Coastline as a combined renderer. Its generic ocean setup uses an infinite mesh and simulation border blending; this finite-lake hook deliberately retains mode 5 and the existing saved simulation bounds.