# Isolated unchanged-cut candidate and geometry cost attribution

Registered before native measurements,2026-11-13. Full C1 remains open:
actual uniform20->10 at10us fails space49.56%dh/76.12%motion; original
50/200us, physical normal flux/traction and physical predictions not achieved.
This small experiment decides whether an exact geometry optimization is worth
integrating; it cannot pass or replace any physical/time/space gate.

Use the four frozen cost_profile_20261113 native inputs, code units and domain.
Existing headers/binaries/solver unchanged. New isolated function keeps original
site/cut/vertex order and intersection arithmetic, and skips allocation only
when EVERY vertex satisfies the original <=0 floating predicate. No radius
screening, neighbor truncation, shape force or tolerance change. Require bitwise
identical vertex coordinates, order and count, plus original partition checks.
This is finite-state evidence; production adoption would additionally require
invalid-input/degeneracy coverage, energy/gradient and actual coupled controls.

Reference for established plane-intersection screening (not a novelty claim):
https://math.lbl.gov/voro%2B%2B/doc/refman/
https://math.lbl.gov/voro%2B%2B/examples/cut_region/
Voro++ also documents radius/region tests; neither is implemented here.

Time reference clipping, candidate clipping, native post-cleanup partition
validation and full reference geometry. Warm all once,5samples each; alternate
reference/candidate clipping order. Timings overlap and cannot be summed.
Record all samples; host load still limits inference. Require median clipping
speedup>=1.25 for BOTH N544 states before considering integration, and inspect
clipping share of full construction. If contribution is small, leave candidate
isolated and redirect effort. This gate is an engineering screen, not science.
4 sequential cases,120s per case,total600s hard cap. No >2h job. Preserve inputs,
outputs,source/binary hashes,logs and any failure. PNG+vectorPDF white red/blue
report, bounded byte-exact raw ZIP, all4 research docs and own public subpage.