Follow-ups to the cleanup review, where noise-level result changes were accepted for simpler code:
- one PICGRID reader, picurv_cli/picgrid.py, now serves the conductor and every generator, and grid staging refuses a non-finite coordinate;
- particles_per_cell, source constants, inlet fluctuation length_scales and time_scale, and the reducer friction value refuse inf and nan; an infinite particles_per_cell is reported instead of ending in an OverflowError;
- Brownian increments come from the shared keyed normal pairs in expression.c: still keyed on (seed, PID, step), but a different realization of the same distribution, so trajectories differ from earlier builds while statistics agree (MSD within 0.11% of the Einstein relation at 50,000 particles, against 0.08% before);
- a recipe ensemble of a dimensionalized post run is no longer unscaled and rescaled; this also fixes homogeneous profiles of an ensemble pairing an ordinary field with a particle-sampled one, whose occupied-state mean was left scaled.