Particle Lagrangian stress history: an inflow cell is refilled from stale neighbours, not from inflow_value #783

`stress_transport="lagrangian"` (`uw.systems.ddt.Lagrangian`) keeps its swarm populated: the constructor installs a `population_control` and every `advection` ends with `repopulate`, refilling a starved cell from the bounded RBF reconstruction of its nearest particles and capping over-full cells. On a closed flow that is complete. On an open flow the value written into a refilled inflow cell is wrong:

• The flavour declares `applies_inflow_value = False`, so a caller cannot say what enters.
• A cell in an inflow arm drains within one transit; the refill reconstructs its particles from whatever old particles are nearest, so the inlet carries a smeared, stale stress instead of the entering state. Every step re-smears it. The integration-point and forward flavours apply `inflow_value` to a departure point that left the domain (Integration-point stress history blows up at the inlet of a plain channel (Wi 1 and 5, Re 100); nodal and grid histories are stable there #745); the particle flavour has no equivalent.

Evidence: the 2026-09-23 domain-of-applicability sweep, cross-slot De 0.3, β 0, dt 0.2, refinement 1. The four mesh flavours complete (eulerian and forward steady at 61 steps; IP ran to the cap; nodal stalled). The Lagrangian cell recorded one DIVERGED_LINEAR_SOLVE, then hung at step 62 (t 12.4, domain length 10, unit inflow) for five hours at full CPU with no output and was killed by hand. The hang itself is not yet diagnosed and may be a separate defect.

Fix: honour `inflow_value` on `Lagrangian`. After the advection (and its repopulation), a particle whose one-step back-trace `x - u dt` leaves the domain entered through an inflow and takes the inflow datum in every history slot, the same test the integration-point flavour applies to a restored departure point.

Branch: `feature/forward-parallel` (no PR yet).