AtmosphereField

Two sources on purpose. Soot a fire produced is the C(soot) species, which keeps its carbon inside elementTotals — a channel carrying mass outside the ledger would make the most valuable assertion in the track quietly wrong every time something smoked. The channel is for a consumer who wants smoke without chemistry behind it: a smoke grenade, a chimney, a dust cloud.

Soot is near-black and a droplet is near-white, so the ratio between them is what colours a plume: §17 reads it, and a wet fire steams white where a rich one smokes black with neither being chosen. Symmetric with sootAt, and for the same two reasons.

Two sources, exactly as soot has. Anything the chemistry condensed is a species — wood tar coming out of a cooling plume is the case this was built for — so its mass stays inside elementTotals and the most valuable assertion in the track keeps covering it. The channel addAerosol writes is for a consumer who wants steam without a boiler behind it: a kettle dressed for a scene, a waterfall's spray, a breath on a cold morning.

The field's own number rather than a second one. §17 gives a smoke particle its velocity from "the cell's advected plus buoyant velocity", and this is the buoyant half — the same function transport uses, so a particle rises at the speed the gas carrying it actually rises at rather than at a rate chosen to look right.

ambientTemperatureExact rather than this.ambient.temperature, and the difference is a determinism bug rather than a rounding nicety. A cell nobody has touched is not stored; one a neighbour realised is. Returning the authored 293.15 for the first and the derived 293.15000000000003 for the second made a cell's reported temperature depend on whether some other parcel had happened to create the chunk first — which put an order dependence into convection, and from there into whether a log ignites on tick N or tick N+1.

Found by world/fingerprint.test.ts: two logs spawned the other way round diverged in the eleventh digit on their very first tick, and nothing about them was different.

What a flammability limit is checked against. Below the lower limit there is not enough fuel to carry a flame; above the upper there is not enough air, which is why a fuel-soaked rag smokes rather than flames until it thins out.

The quantity visibilityAt is derived from, and the one a consumer attenuating a light along a ray actually needs: metres of visibility is one answer to it and exp(−σ·d) is another. Exporting only the first would have every consumer inverting Koschmieder to get back here.

Soot and droplets both. It counted soot alone, so a smoulder's smoke, which is droplets, hid nothing and was drawn at an alpha of about 1e-29 by emitSmoke: the pale half of smoke was in the field and nowhere on screen.

Koschmieder over the smoke's extinction, soot and droplets. Clamped at ten kilometres, past which the answer stops being about the smoke.

The reading that decides whether anything dries. §8.1's evaporation is driven by the vapour pressure deficit, so washing dries in wind and nothing dries in fog — and the same number is why a fire dries a room out while removing no water at all: warm air could hold more, so the same vapour is a smaller fraction of saturation.

Saturation is tabulated, per §7's rule, because it is exp of a rational function of temperature and this is a query a script may make every frame. 256 knots over 200-500 K, where linear interpolation is under a tenth of a percent.

Every transfer is flux-form and conservative — computed once for a face and applied to both sides — which is a departure from §9 worth recording. The design chose semi-Lagrangian advection for unconditional stability; semi-Lagrangian is not conservative, and conservation is the invariant this whole track is asserted on. Upwind flux has a CFL limit instead, and a CFL limit is handled here the same way stiffness is handled in the reaction solver: by sub-stepping. That is the trade taken deliberately.

Deltas are accumulated and applied afterwards, so no cell sees a neighbour that an earlier cell in the sweep already changed. A Gauss-Seidel sweep would give an answer that depended on chunk allocation order, which is a determinism hole with no symptom until two runs allocate differently.

windX and windZ come from the one WindField, sampled once per frame by the caller.

Live chunks alone do not balance, and the reason is not transport: creating a chunk draws a chunk of air in from outside, which is matter appearing in the live set. The ledger records it, so elementTotals + ventedElementTotals is invariant under chunk creation as well as under everything else — and a conservation assertion over a growing plume is possible at all.