Every cell did an atomicMax on the same word; a workgroup reduction first
leaves one per 256 cells. Same result, less contention on a single
address.
Part of #529.
The CPU stays the default, so every existing result and fixture is
bit-identical; "gpu" steps on the device FIELDCRAFT_GPU picks through the
registered backend, and is an error when the build has none. The desktop
app and the CLI (hence the MCP server) register fieldcraft-lbm-gpu at
start; the summary names what stepped. Hostile values are errors.
Part of #529.
A new L7 crate: the CPU scheme's collision (TRT with Guo forcing), the
per-link streaming rules (bounce-back with Ladd's term, slip reflection,
open placeholders), the Zou–He boundaries and the momentum exchange per
force group as wgpu compute shaders, stepping the flow fieldcraft_lbm's
Flow prepared, through its Backend trait: populations stay on the device,
a chunk of steps is one submission, and only frames, forces and the
stability check come back. Populations are stored as deviations from
their weights, which keeps single precision honest: the verification
cases (Poiseuille with Λ = 3/16, Couette with wall forces, Taylor–Green,
a sphere's drag, a Zou–He channel) agree with the CPU to 2–5e-6 in
velocity, 1e-6 in density and 5e-5 in force, with those tolerances stated
in the tests. The rules are applied one thread per link rather than in a
per-cell loop, which lavapipe's JIT miscompiles (and which diverges on
real GPUs anyway). Native only for now; the browser registers nothing.
Part of #529.
run() drives a Backend in chunks of steps (to the next frame, history sample
or progress check) instead of the CPU Flow directly, so a device backend
registered by an application can step the same prepared flow: Flow keeps
building the grid, links and open cells and exposes them read-only, and
implements the trait itself. solve_fluid_on picks the CPU or the registered
device backend; solve_fluid stays the CPU and is unchanged in its results.
Part of #529.