rdb_ocean_min_thickness Module

Conservative minimum-layer-thickness adjustment for the isopycnal (VCOORD_LAGRANGIAN) ocean path.

Motivation: in the remap-free Lagrangian coordinate an eddy can displace an interface so far that a layer thins to ~0 (“outcropping”), and then u = hu/h blows up. The existing angstrom_h floor lifts a sub-floor layer with max(h_new, angstrom_h), which INJECTS mass (non-conservative: total column thickness, SSH, and tracer mass all drift up on every floored step) and the injected volume shapes the eddy field.

This module instead performs a CONSERVATIVE adjustment: when a layer thins below the floor, the deficit is borrowed from the surplus layers of the SAME water column — the interface is moved, no mass is created. Per water column the following are conserved to round-off: * total thickness Sigma_k h_k (SSH unchanged) * momentum Sigma_k h_face_k . u_face_k (per C-grid face) * every tracer mass Sigma_k h_k . Tr_k

The adjustment is a STRICT no-op on any column where all layers already meet the floor: those columns (and their faces) are left byte-unchanged so the isopycnal interface structure — and hence the baroclinic mode we are modelling — is never pinned. This is the load-bearing constraint that distinguishes this from a z*/sigma remap.

Implementation: reuse the conservative per-column remap_column primitive (the same engine the ALE remap uses) with a “floor-only” target column — each sub-floor layer inflated to the floor, the excess drawn conservatively from the surplus layers so the column total is preserved.

COST STRUCTURE (the 31%-of-runtime restructure, 2026-07-24): the borrow is a no-op on every non-grounded column, but the original implementation still paid ~8 full-field passes per call on a healthy domain (unconditional target build + copy-back, per-tracer concentration divisions before the gate, and per-face NEIGHBOUR COLUMN re-scans with k-strided access). Now: ONE coalesced pass builds a 2D grounded mask + a global count; zero grounded columns ⇒ the whole call returns after that single read; otherwise every kernel consults the mask (2 loads) and only grounded columns / active faces do work — and reads fall back to h_old on non-grounded neighbours, where h_new ≡ h_old by construction. Byte-identical results in all cases.