peclet-flow 1.4.0

Added

• Cut-cell scalar transport with solids (opt-in: add_scalar(name, D, cutcell=True)). A
  transported scalar sees the immersed bodies as cut cells with a probe-flux wall condition, second
  order for Neumann, Dirichlet and Robin walls (set_scalar_wall), conjugate transport into the
  solid (set_scalar_solid, get_scalar_solid), sources (set_scalar_source), a mean-gradient
  closure for periodic beds (set_scalar_mean_gradient, diagnostics.scalar_mean_flux), and a
  steady solve (solve_scalar_steady) beside the transient advance; set_scalar_residual_tolerance
  (the velocity solver's criterion, one spelling), diagnostics.set_scalar_max_iterations(max_iter);
  per-body wall fluxes diagnostics.scalar_wall_flux, budgets and
  solver statistics diagnostics.scalar_budget / scalar_census, the cut-cell record
  diagnostics.scalar_geometry / scalar_facets. Conservation is exact in the discretization; the
  solver residual is reported, never fixed up. Staggered grids and the collocated cut-cell
  projection; refused by name: the collocated 'ghost' scheme, collocated open faces, porous
  continuity, moving scenes. Legacy scalars (cutcell=False, the default) are unchanged.
• diagnostics.set_pressure_vcycle_precision('fp64' | 'fp32' | 'auto') and
  diagnostics.pressure_vcycle_precision — an FP32 V-cycle below the exact FP64 PCG/FCG, with an
  FP64 restart (latched for the run) on a cap or breakdown; single rank and distributed.
• diagnostics.set_pressure_initial_guess('zero' | 'projected') (opt-in; default 'zero'):
  A-norm projection onto the last one or two pressure solutions.
• diagnostics.set_pressure_prefactor_hoist(bool), diagnostics.num_pressure_prefactors,
  diagnostics.num_pressure_chebyshev_restarts (developer tier).
• With np > 1, init_mpi prints one line on rank 0 naming ranks whose OpenMP threads straddle L3
  (CCD) domains (no result effect). initMpi(dec) refuses, on every rank, a decomposition the ranks
  do not share.

Changed (named numerics changes)

• The GPU pressure V-cycle defaults to 'auto': FP32 where eligible on device backends, at
  every np, with the FP64 restart; host 'auto' is 'fp64', so host results are bitwise unchanged.
  Device fields move in the last digits (~1e-13 relative); 66 probe configurations 0 violations,
  2000 steps 0 restarts. ~5 % per step on one H100 (16.9 → 16.1 ms), −28 % projection on an
  RTX 5080, never slower at np > 1. 'fp64' restores the previous solve.
• The default decomposition (set_decomposition(0)) is the AUTO policy: the legacy aligned ORB
  where it is exactly balanced (bit-identical there), otherwise a balance-aware coarse-first
  partition (core SplitAxis::Balanced), otherwise the better of legacy and the balance-aware
  aligned ORB; it never cuts an axis the legacy ORB leaves whole. np > 1 partitions change where
  the legacy split is inexact (bubble column np 3/12/24): fields within 5.9e-14 of the legacy
  partition, iterations identical; the bubble column at 24×1 was 2:1 imbalanced under legacy.
  set_decomposition(1) is the legacy ORB; explicit levels ≥ 2 are unchanged (2394/2394 published
  partitions byte-identical). The policy keeps legacy's set of split axes unless legacy's max/mean
  load is ≥ 5/4; this guard reverts 14 census configurations to the legacy partition (128×96×64
  at np 12 among them). The box-shell launch of the auto 12×2 column no longer runs serially on one
  thread (6.2 ms/step on genoa); bitwise.
• Host backends use the 'direct' pressure bottom where eligible (single rank: FP32
  block-tridiagonal factor + FP64 FCG, replacing GraphAMG; distributed: solved redundantly on every
  rank from the allgathered bottom). 50-step difference 3.3e-14 single rank, 2.3–3.6e-14 at np > 1,
  pressure iterations identical; set_pressure_bottom_solver('algebraic') restores GraphAMG.
  Workstation 8×3: 61.4 → 53.6 ms/step.
• VoF tier-3 curvature fallback (core 1.5.0's moment kernels): a recorded round-off change, 50-step
  max relative 2.7e-14 host / 2.5e-14 CUDA, iterations identical; tier 3 3.55 → 0.58 ms on an
  RTX 5080.
• Host pressure-MG sum reductions run in pencil order; the Chebyshev driver re-estimates its spectral
  bounds from the previous estimate after a variable-density rebuild (cold re-estimate on a guard,
  counted); VoF block statistics are reduced in one launch per step (device: fixed 512-cell tiles) —
  diagnostics order only.
• Scalar and phase-change energy inputs are physical under an armed extent (a fix of the
  physical-units contract): add_scalar's diffusivity, mass fluxes, latent heat, conductivity,
  ρc_p and the closures on ρ/μ/forces are converted at the boundary; the getters report physical.
  Scripts that passed cell-unit values to a solver built with extent= change result.
• VoF blocks under a physical domain (Solver(cells, extent=…) with enable_vof_blocks): the
  step handed its internal dt to the physical-unit block advector, which converted it a second
  time (the first step hit the Weymouth–Yue cap), and the block CSF kept a σ from before the first
  set_dt (1/t_ref² too strong). Both fixed; cell-unit runs are bitwise unchanged (ctest
  vof_blocks_physical_units).
• Open boundaries keep their high-side face plane. Ghost fills of the MAC face fields wrapped
  every axis, so a non-periodic high-side face (outflow or inflow) carried the low boundary's
  face; the collocated projection and the collocated VoF bridge exposed it, and legacy scalars
  read the wrapped plane. Periodic axes are untouched (ctest open_face_plane, np 1–3).
• Scalar inflow. A scalar's ghosts are filled before the old state is frozen, so the first
  step after set_field (or a coupling or VoF write) no longer upwinds from stale ghosts; at a
  Dirichlet face that is also a velocity inflow the advection upwinds the prescribed value, not the
  Dirichlet reflection (a plug's inventory was U t + 0.5 cells; now U t to 6e-14, ctest
  scalar_inflow). Results change for scalars written before the first step (state hash
  scalar). Trade: on a smooth inflow profile the first interior face carries an O(h) error
  (gh/6 coded limiter, gh/4 second-order upwind).

Changed (performance, bitwise)

• Distributed host step: the pressure-MG stencil exchanges are faces-only with rank-local periodic
  axes read in-kernel; core 1.6.0's halo transport (persistent requests, one pack launch, on-node
  shared memory: 10.9–12.5 → 1.1–1.7 µs per exchange); fused host Krylov reductions; host VoF
  kernels and momentum ghost fills in fewer launches.
• Device step: no global synchronization inside step() (ctest device_sync_budget), GPU-resident
  coarse tail with the bottom factor on a side execution instance.

Known limitations

• Cut-cell scalars: particles in contact do not converge at R/h 8–32 (no contact model yet);
  collocated open faces are refused.
• Distributed host runs: threads per rank must divide the cores of an L3 domain; 8×3-type layouts
  on 8-core CCDs are slow (+17.6 ms kernel time on genoa), whatever the partition.

Part of the peclet family 1.5.0 (pins core v1.7.0).