# Actual uniform5um startup and one-step cost qualification

Registered2026-11-13before execution. Preserve original full C1 scope; actual
20->10space at10us still fails49.56%dh/76.12%motion. Full curved-energy
directional response supports testing resolution, not proven cause. No new
physics/kinetics/shape forces, no tolerance change or empirical fitting.

Run original tf_curved_edge_bench binary SHA256
76e49ed253e2341df37f93d02803272d020b8500ed9265da6f7d0ec124b6c5c3,
no candidate dispatch. Same SI parameters and preparation as global_space:
uniform5um,320x160um,physicalcut240um,radius500um,SDS6mM,insoluble,
initial6um,no gravity/DLVO,constant-activity variational preparation,trace
released,bath fixed height/potential. ExpectedN2112(2048bulk+32trace+32bath).

Two NEW one-step actual coupled cases:dt.0625/.03125us,T=dt,save_stride1,
observer on,refine_start200um(ignored by uniform native layout).
No frozen successful trajectory rerun. Verify original24native checks via
hashes and exact log counts by name(observer6,preparation4,curved14), not new
test execution. Frozen original source subset and original binaries must
still match global_space protocol. Extra new standalone diagnostic files do
not replace any original source or change its behavior.

Require complete,failures0,N2112,all accepted force ratios<=1.000001,
global/regional V/S<=1e-10,energy budget<=1e-7,prep pressure<=1e-8.
Native guard timeout600s each,total native1200s. Preserve partial/timeouts,
do not restart timed-out native cases. If first guard fails, skip second.
Independent saved-step scalar and mechanical SI/GEOS audits on immutable
copies, unchanged original formulas/tolerances, one external process cap300s
per terminal passing case. Preserve failed/partial audit attempts, do not
silently retry. No>2h task; native+audits capped1800s before presentation.

Record driver wall_s and native summary.wall_s separately. The native timer
starts after preparation/initial exports; driver-minus-native includes setup,
preparation and IO, not a pure preparation benchmark. Single-step cost is a
budget observation, NOT a controlled scaling law or guaranteed long runtime.
Engineering projections for each10/50/200us atmatchingdt:
1.3*(outside_native_timer + required_steps*native_timer), with explicit host
load/nonlinear-iteration uncertainty. Do not auto-launch longer matrix from
this plan. Each actual extension requires its own frozen budget/guards and,
if estimated>2h, a half-hour monitor set BEFORE launch.

One-step cases end atdifferent physical times; they are NOT a time refinement
pair and do not establish10us time/space acceptance. Original10/50/200us,
actual global convergence, independent normal flux/traction, exchange
coefficient/new thin-element ratio and experimental fringe/heldout prediction
remain required. Earlier4regular area-FD failures and source geometry-candidate
11byte gate failure(timing-only) retain their original states.

All raw native bytes,logs,full source/input/binary hashes,projections,audit
coverage,PNG+vectorPDF white Nature red/blue report archived; update all4
research entry docs and own VariDrainage subpage. Full C1 remains open.