Scope: approved A, coarse slot/reservoir with an internal fixed patch. Only the three files in incremental.patch changed; all other baseline files retain their hashes. No dependencies installed, no DynamisLab access, no commit/push.
API: FixedPatchStepper(cfg, bounds, context=ctx, coarse_boundary="slot_reservoir"). PeriodicPatchStepper remains an alias with periodic default.
Order: coarse pull -> coalescence -> existing planar BC -> collision. Fine retains four-cell halo and no physical BC. Fine coarse-only settings are neutralized. Wall-aware patch requires y0>=4, y1<=ny-4; periodic margin2 unchanged. Covered coarse, outside-wall rows and fine halo are not public leaves.
References consulted before implementation/design: ref/lbmpy/src/lbmpy/relaxationrates.py:61 (acoustic tau scaling); ref/palabos/src/multiGrid/multiGridGenerator2D.hh:190 (two fine substeps); existing boundary/planar_velocity.cuh closures reused unchanged.
Validation: 13 CPU tests; prescribed 32x24, bounds(12,4,20,10), W8/A1e-3, tau0.8/1.1, 8 coarse/16 fine steps against independent scalar FP64 G-to-G reference; one existing 17x13 margin2 rectangular periodic step. No old eight-step matrix rerun. Both GPU cases in one job after fresh GPU0 UUID/occupancy/RAM precheck, exact command and resource monitor in results.json.
Population tolerance fixed in advance 3e-6, BC tolerance 2e-6; measured maxima 1.78241457e-7 and 2.21654773e-7. Periodic max error 1.14050047e-7. Raw arrays/NaN masks, eight corner links, actual fine generated macros and public leaves checked. No active NaN hidden by masks.
Open-domain discrete leaf mass change +0.0309633738; final summed bottom/top rho*v 0.00426894613/0.000960694626. This is reporting, not an assertion of constant mass or proof of global discrete flux balance. Coarse owns the nozzle; no wall-touching fine patch, steady jet or convergence claim.
Two read-only agent reviews found no blocker; oracle uses independent scalar transport/collision and the same analytic BC model, plus independent macro checks.
Baseline provenance: production snapshots taken before edits. Test edits began before snapshot; test baseline reconstructed from delegate's exact edits and matches prior example manifest SHA256 d2df924dfd5267e732036c6df27bb8809f59f1d371da75f506c67aad5997f322. Prior evidence retained unchanged.
Resources: live RAM/VRAM reserves, no fixed 256MiB RSS cap; combined temporary budget 2,000,000,000 bytes. Peak monitored temporary bytes 1,548,161; GPU job exit0. All task-owned compilation scratch and baseline copies removed after diff; monitor is sampled, not an OS hard limit.
