Second round of review follow-ups, and synthetic inflow fluctuations promoted on a channel benchmark:
- CSV records find a run's first write to each file themselves: output-cadence records mark a continuation wherever their first row falls, the Poisson history keeps the initial condition's step-0 rows, and the Newton-Krylov and Poisson histories are no longer re-truncated within the first step;
- an inlet's particles_per_cell becomes a density through the volume Volume seeding measures, so the two agree on curved grids and at any rank layout;
- named sources that refer back to themselves are refused with a message, and a sum named under sources: is projected by default as an inline one is;
- the wall-model routines are untimed per-cell kernels, at their measured call rates;
- fluctuations on a prescribed_velocity inlet is no longer experimental: synthetic-inflow-channel-2026-11-08 recovers a Re_tau 203 channel's skin friction within 5% from 15.1 half-heights, where a mean-only inlet stays laminar.