# Retained-face actual coupled single-step controls

Preregister before new native execution. Require the7-input retained-face
geometry admission and24new curved/preparation/observer fixture checks.
Compile an opt-in wrapper selecting retained geometry and the SAME local
Hessian energy assembly, CHOLMOD/H1 backend, original benchmark, parameters
and original tolerances. Do not modify default/frozen original source/binaries.

Execute exactly3NEW candidate trajectories: uniform5um dt=.0625us one step,
uniform5um dt=.03125us one step, uniform10um dt=.0625us one step. All: radius
500um,amplitude0,SDS6mM insoluble,variational preparation,save_stride1,
refine_start200um,observer true. Reuse immutable successful local_assembly
outputs as reference; never rerun original trajectories. For the10um reference
use ONLY its saved first step from the old160-step run.

Native gates unchanged: one accepted requested step, failures0, expectedN,
force/target<=1.000001, maxglobal/regionalV/S<=1e-10, energy excess<=1e-7,
preparation pressure<=1e-8. Independent original SI/GEOS scalar and mechanical
formulas audit all saved candidate steps. Adapt paths and24reused-check metadata
only. Preserve any native/equivalence/audit failure; stop new cases on first
native/equivalence failure. No automatic retry or tolerance tuning.

For output comparison, retain UNMODIFIED local_assembly/equivalence.py column
norms and metadata comparisons. The explicit adapter selects step1 in
states/faces on BOTH sides, and compares mesh/preparation JSON only. Exclude
whole-trajectory summary JSON because the10um reference spans160steps; native
candidate summary is separately checked by original native gates. Initial
states/faces remain complete. Thus this is first-step consistency, not whole
trajectory equivalence or10us time/space acceptance. Velocity/face-flux
tolerances remain1e-8; other continuous columns1e-10, x/y normalized bydx;
discrete structure exact, unchanged from original numerical-equivalence rule.

Configure/build caps300/600s; native per-case240s,whole<=720s; independent
audit per-case180s. No >2h computation. Export actual saved particle fields,
native checks, independent audit seals and raw bytes with Nature red/blue
PNG/vectorPDF and white standalone HTML. No fitted rate or C2 change.

Only if all3controls and independent audits pass, consider a new SAME-candidate
time pair at10us. Do not combine an old-geometry coarse step with new-geometry
fine step as a convergence pair. Any >2h batch needs a verified half-hour monitor
BEFORE launch. Original failed pair,10/50/200us,physical flux/traction and fullC1
remain open. These3one-step controls do not exercise actual rejected78state.