Close the heat-transfer and electrical gaps against ShipSIM

[pankgeorg]

Closes the remaining ShipSIM components that had been left out, plus three bugs the work exposed. Every claim below is checked by an assertion in one of the two validation scripts.

Heat transfer

The four ConvectionFactors. Upstream's Internal.ConvectionBaseClass is
ThermalComponents.Interfaces.ConvectiveElement1D — same solid/fluid ports, same ΔT,
same sign convention — so Thermal.ConvectionBaseClass extends the library interface and
adds only A, h, Tavg and Q_flow = h·A·ΔT. The porting notes had blamed
BuildingsHeatTransfer not resolving; it was never the blocker. Upstream gives its four
factors two different shapes (two extend the base class, two wrap a Convection and drive
its Gc); all four extend the base class here, so they report h and A the same way —
which is what ShipCompartment then consumes.

ConvRadSunWall rebuilt as a composition. The previous port flattened it into one
algebraic flux and took h_c as an input. It now composes what upstream composes: two
BodyRadiation paths split by view factor, ExternalConvection, IrradiationOnPlane,
SunScreen and two PrescribedHeatFlows. The flattened version had radiated the whole
surface to the sky, overstating night-time cooling on a vertical wall by a factor of two,
and had no transmission port, no sun-screen wiring and no sun-height gate.

ShipCompartment is MultiportLumpedRoom with its constant U_inner/U_outer
replaced by the two convection factors. A building's outside film coefficient is near
enough constant; a ship carries its own weather. Alongside, the two models agree (the
fixed U_outer = 5 really is the still-air 5.36); at sea the fixed-U model runs 2.75 K
hot behind insulation, and 9.30 K on bare steel where the films are nearly the whole
resistance.

EnvironmentHeatTransfer derives cloud cover from the irradiance shortfall against
clear sky — a clocked moving average with a night latch, where upstream uses
Mean(f=1/1800) and a when IsDay == false — plus sky temperature, wind vector and a
Magnus dew point (upstream's linear form asserts it is only valid above 50 % RH).

Electrical and fatigue

RandomStart is upstream's when time >= pre(NextStep) scheduler written in the
clocked sublanguage on DiscreteComponents.UniformNoise — seeded and reproducible. The
note in OnOffConsumer saying this could not be expressed in Dyad predated
DiscreteComponents.

ElectricalLoad is the EPLA that MiscSamples.Balance performs, with the consumer
list as parameter arrays rather than 177 hand-written instances. Six consumers over six
hours: bus load 0–67 kW across 51 distinct levels, 24.9 kW average against 79.1 kW
installed, diversity factor 0.314.

Rainflow as Julia, not a component. src/Rainflow.jl counts cycles (ASTM E1049
three-point), bins the matrix and accumulates Miner damage over the turning points
EventPeakSampler already emits. Upstream runs it inside the model on an algorithm
section with two stacks and while loops. The clocked layer could express a bounded
version, but counting cycles is post-processing over a series: it has no state the rest of
the model can see, it needs an unbounded stack to be exact, and clocked array state does
not cross-compile (SynchToolkit #212), so a model carrying one could not be deployed.

Three bugs this exposed

• sun_vector_world had the sun in the northern sky. It negated the east and up
  components to turn direction-to-sun into ray direction, but not the north one. A
  south-facing wall was lit at dawn and dusk and dark at noon. Latent since the solar port:
  IrradianceOnPanel uses only the vertical component and nothing else consumed
  SunVector_y. The two solstice analyses are the regression check — in winter the south
  wall takes the whole gain and the north none; in summer that inverts, which is the real
  behaviour of a vertical wall at 43°N.
• ExternalConvection was not differentiable at rest. sqrt(u_x²+u_y²+u_z²) has no
  gradient at the origin (0/0) and |u|^0.78 an unbounded one; a max(|u|, tiny) floor
  bounds neither. Either puts a NaN or a 2400 W/(m²·K) per m/s entry in the Jacobian at
  t = 0. Between two ideal temperature sources neither shows — which is why the §1 suite
  passed — but with a wall face that has no capacitance the integrator drives dt below
  floating-point epsilon and reports only "unstable". A steady wind of any speed hides
  it, zero included; it is passing through zero that fails, which is what a ship leaving
  harbour does.
• A ZeroOrderHold collapses everything downstream of it, not just itself. The known
  caveat is that its continuous output reads back as its final value under symbolic
  indexing. It also collapses anything algebraically derived from it. An ElectricalLoad
  bank read back as a flat line, every consumer that happened to be off at the end
  reporting zero for the whole run, while the model had integrated all six correctly. Fixed
  by putting a state in the path in both places it bit — a switching lag in OnOffConsumer,
  a one-tick lag on the cloud ratio.

Library findings, recorded in PORTING_NOTES.md

• HVACComponents 0.3.0 will not initialise a MultiportVolume whose moist-air fluid
  boundary is below about 15 °C. Minimal reproducer with no component of ours in it; the
  media's own property functions are exact down to −5 °C, so it is in the assembled DAE.
  Only fluid ports are affected, so the deckhouse still runs its 0 °C winter case on a
  preheated supply — which is what a ship does anyway.
• ThreeTempWall1D.m_wall and OneTempWall1D.m_wall are documented as the total mass
  across N surfaces but each surface's balance divides by the whole of it.

Validation

scripts/validate_heattransfer.jl — 93 assertions. scripts/validate_machinery.jl — 43.
Both green on a clean rm -rf generated && dyad compile. Correlations are recomputed from
the literature in the test rather than read off the port; the wall's radiation split is
checked against the view factors on the solver's own saved steps; rainflow is checked
against hand-worked sequences.

Not ported, and why, is in PORTING_NOTES.md: TriggerConsumer/StartGenerator (variants
of RandomStart, writable the same way), SunIrradianceMultibody, SubComponents.Ikeda
(a stub upstream) and VariableTranslation.

🤖 Generated with Claude Code