feat(num): N3 selectable scale-floor clamp mode (HardJava | SmoothBarrier)

The hyper-ideal vertex scale b is floored to keep the geometry valid.  The
original clamp `b<0 → 0.01` mirrors the Java oracle but is only C⁰ (in fact
value-discontinuous at b=0): a Newton step crossing the feasibility boundary
hits a kink that can stall convergence (numerical-stability audit N3).

Rather than replace the Java-faithful behaviour (which would break the golden
parity tests), make the floor a selectable mode so BOTH the Java standpoint
and the clean mathematics are available:

- HyperIdealScaleClamp::HardJava (DEFAULT) — the original snap, bit-for-bit
  faithful to HyperIdealFunctional.java → all parity tests unchanged.
- HyperIdealScaleClamp::SmoothBarrier — C¹ softplus floor
  b ↦ floor + softplus_β(b−floor), β = HYPER_IDEAL_SCALE_SHARPNESS (=100);
  ≈ identity away from the floor, smooth across b=0.  Opt-in.

clamp_hyper_ideal_scale centralises the logic (also folds in the N4 nachzügler:
compute_face_angles used a bare 0.01).  The mode threads with a defaulted
trailing parameter through compute_face_angles, face_angles_from_local_dofs,
evaluate_hyper_ideal, the four hyper_ideal_hessian* variants and
newton_hyper_ideal — so every existing call site keeps HardJava behaviour.

Tests (+4): clamp-function C¹/floor/identity contract, mode-equivalence away
from the boundary, and end-to-end SmoothBarrier convergence to the same Java
golden vector (LawsonHyperIdeal).  296/296 CGAL tests pass.
Documented in doc/math/geometry-modes.md.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

code/include/constants.hpp

@@ -29,6 +29,13 @@ constexpr double LOG_EDGE_LENGTH_FLOOR = -30.0;
 /// Rationale: ensures b > 0 maintains the Lorentzian model's causal structure.
 constexpr double HYPER_IDEAL_SCALE_FLOOR = 0.01;
 
+/// Sharpness β of the optional C¹ smooth-barrier scale floor (N3 audit).
+/// Only used when the hyper-ideal clamp mode is `SmoothBarrier`; larger β
+/// makes the softplus transition tighter around `HYPER_IDEAL_SCALE_FLOOR`
+/// (β·floor ≈ 1, so β ≈ 100 keeps the barrier ≈ identity for b ≳ 0.05 while
+/// staying C¹ across b = 0).  See `clamp_hyper_ideal_scale`.
+constexpr double HYPER_IDEAL_SCALE_SHARPNESS = 100.0;
+
 /// Domain guard for asin in spherical_l (spherical_geometry:34).
 /// Clamps exp(λ/2) to slightly below 1.0 so asin stays in [−π/2, π/2].
 /// Rationale: asin(x) is undefined for |x| > 1; this prevents IEEE Inf/NaN.

code/include/hyper_ideal_functional.hpp

@@ -158,6 +158,55 @@ static inline std::size_t hidx(Halfedge_index h) { return halfedge_to_index(h);
 // HyperIdealFunctional.java's defensive behaviour (lines 122-127 of the
 // Java original); this keeps the FD perturbation regime well-defined.
 
+// ── Scale-floor clamp mode (N3 audit) ─────────────────────────────────────────
+//
+// The vertex scale b must stay positive for the hyper-ideal geometry to be
+// valid.  Two ways to enforce that are offered, selectable per call:
+//
+//   • HardJava (DEFAULT) — the original `b < 0 → HYPER_IDEAL_SCALE_FLOOR` snap.
+//     Bit-for-bit faithful to HyperIdealFunctional.java, so the Java
+//     golden-oracle parity tests hold exactly.  It is only C⁰ at b = 0: a
+//     Newton step that crosses the feasibility boundary sees a kink, which can
+//     stall convergence (numerical-stability audit N3).  This is the
+//     production default precisely to preserve parity.
+//
+//   • SmoothBarrier — a C¹ softplus floor
+//         b ↦ floor + softplus_β(b − floor),   softplus_β(x) = log(1+e^{βx})/β
+//     which is ≈ b for b well above the floor (to machine precision once
+//     β·(b−floor) ≳ 35) and decays smoothly to `floor` as b → −∞, with a
+//     continuous derivative everywhere.  This removes the N3 kink and is the
+//     mathematically clean choice, but it perturbs values near the boundary
+//     and therefore deviates from the Java oracle — opt in when robustness of
+//     a boundary-crossing solve matters more than strict parity.
+//
+// Both modes share the same floor (`HYPER_IDEAL_SCALE_FLOOR`); SmoothBarrier
+// additionally uses `HYPER_IDEAL_SCALE_SHARPNESS` (β).  The `a < 0 → 0` edge
+// clamp is unaffected (a = 0 is a genuine geometric floor, not flagged by N3).
+enum class HyperIdealScaleClamp {
+    HardJava,       ///< b<0 → floor.  C⁰, Java-parity-faithful (default).
+    SmoothBarrier   ///< C¹ softplus floor; clean but deviates from Java near b=0.
+};
+
+/// Apply the vertex-scale floor to `b` under the chosen clamp `mode`.
+/// `HardJava` reproduces the original snap; `SmoothBarrier` is the C¹
+/// softplus floor (see `HyperIdealScaleClamp`).  `variable` mirrors the
+/// call-site guard (only clamp DOFs that are actually free / variable).
+inline double clamp_hyper_ideal_scale(double b, bool variable,
+                                      HyperIdealScaleClamp mode) noexcept
+{
+    if (!variable) return b;
+    if (mode == HyperIdealScaleClamp::HardJava)
+        return b < 0.0 ? HYPER_IDEAL_SCALE_FLOOR : b;
+
+    // SmoothBarrier: floor + softplus_β(b − floor), evaluated stably.
+    const double beta = HYPER_IDEAL_SCALE_SHARPNESS;
+    const double x    = beta * (b - HYPER_IDEAL_SCALE_FLOOR);
+    // softplus_β(x) = log1p(e^{βx})/β, with the standard large-x guard
+    // (for x ≳ 35, log1p(e^x) == x to double precision) to avoid overflow.
+    const double softplus = (x > 35.0) ? x : std::log1p(std::exp(x));
+    return HYPER_IDEAL_SCALE_FLOOR + softplus / beta;
+}
+
 /// Six per-face angle outputs computed from local DOFs (see
 /// `face_angles_from_local_dofs`).  Used by the block-FD Hessian.
 struct FaceAngleOutputs {
@@ -176,15 +225,16 @@ struct FaceAngleOutputs {
 inline FaceAngleOutputs face_angles_from_local_dofs(
     double b1, double b2, double b3,
     double a12, double a23, double a31,
-    bool   v1b, bool v2b, bool v3b)
+    bool   v1b, bool v2b, bool v3b,
+    HyperIdealScaleClamp clamp = HyperIdealScaleClamp::HardJava)
 {
     // Same defensive clamps as compute_face_angles.
     if (v1b && v2b && a12 < 0.0) a12 = 0.0;
     if (v2b && v3b && a23 < 0.0) a23 = 0.0;
     if (v3b && v1b && a31 < 0.0) a31 = 0.0;
-    if (v1b && b1  < 0.0) b1 = HYPER_IDEAL_SCALE_FLOOR;
+    b1 = clamp_hyper_ideal_scale(b1, v1b, clamp);
-    if (v2b && b2  < 0.0) b2 = HYPER_IDEAL_SCALE_FLOOR;
+    b2 = clamp_hyper_ideal_scale(b2, v2b, clamp);
-    if (v3b && b3  < 0.0) b3 = HYPER_IDEAL_SCALE_FLOOR;
+    b3 = clamp_hyper_ideal_scale(b3, v3b, clamp);
 
     double l12 = lij(b1, b2, a12, v1b, v2b);
     double l23 = lij(b2, b3, a23, v2b, v3b);
@@ -247,7 +297,8 @@ static FaceAngles compute_face_angles(
     const ConformalMesh&       mesh,
     Face_index                 f,
     const std::vector<double>& x,
-    const HyperIdealMaps&      m)
+    const HyperIdealMaps&      m,
+    HyperIdealScaleClamp       clamp = HyperIdealScaleClamp::HardJava)
 {
     Halfedge_index h0 = mesh.halfedge(f);
     Halfedge_index h1 = mesh.next(h0);
@@ -277,9 +328,9 @@ static FaceAngles compute_face_angles(
     if (fa.v1b && fa.v2b && fa.a12 < 0.0) fa.a12 = 0.0;
     if (fa.v2b && fa.v3b && fa.a23 < 0.0) fa.a23 = 0.0;
     if (fa.v3b && fa.v1b && fa.a31 < 0.0) fa.a31 = 0.0;
-    if (fa.v1b && fa.b1  < 0.0)           fa.b1  = 0.01;
+    fa.b1 = clamp_hyper_ideal_scale(fa.b1, fa.v1b, clamp);
-    if (fa.v2b && fa.b2  < 0.0)           fa.b2  = 0.01;
+    fa.b2 = clamp_hyper_ideal_scale(fa.b2, fa.v2b, clamp);
-    if (fa.v3b && fa.b3  < 0.0)           fa.b3  = 0.01;
+    fa.b3 = clamp_hyper_ideal_scale(fa.b3, fa.v3b, clamp);
 
     double l12 = lij(fa.b1, fa.b2, fa.a12, fa.v1b, fa.v2b);
     double l23 = lij(fa.b2, fa.b3, fa.a23, fa.v2b, fa.v3b);
@@ -391,7 +442,8 @@ inline HyperIdealResult evaluate_hyper_ideal(
     const std::vector<double>& x,
     const HyperIdealMaps&      m,
     bool                       need_energy   = true,
-    bool                       need_gradient = true)
+    bool                       need_gradient = true,
+    HyperIdealScaleClamp       clamp = HyperIdealScaleClamp::HardJava)
 {
     HyperIdealResult res;
 
@@ -407,7 +459,7 @@ inline HyperIdealResult evaluate_hyper_ideal(
         Halfedge_index h1 = mesh.next(h0);
         Halfedge_index h2 = mesh.next(h1);
 
-        FaceAngles fa = compute_face_angles(mesh, f, x, m);
+        FaceAngles fa = compute_face_angles(mesh, f, x, m, clamp);
 
         // Store computed angles into temporary arrays.
         // h_alpha[h] = α for the edge of h in this face.

code/include/hyper_ideal_hessian.hpp

@@ -64,7 +64,8 @@ inline Eigen::SparseMatrix<double> hyper_ideal_hessian(
     ConformalMesh&             mesh,
     const std::vector<double>& x,
     const HyperIdealMaps&      m,
-    double                     eps = 1e-5)
+    double                     eps   = 1e-5,
+    HyperIdealScaleClamp       clamp = HyperIdealScaleClamp::HardJava)
 {
     const int n = hyper_ideal_dimension(mesh, m);
     std::vector<Eigen::Triplet<double>> trips;
@@ -77,8 +78,8 @@ inline Eigen::SparseMatrix<double> hyper_ideal_hessian(
         xp[sj] = x[sj] + eps;
         xm[sj] = x[sj] - eps;
 
-        auto Gp = evaluate_hyper_ideal(mesh, xp, m, /*energy=*/false).gradient;
+        auto Gp = evaluate_hyper_ideal(mesh, xp, m, /*energy=*/false, /*grad=*/true, clamp).gradient;
-        auto Gm = evaluate_hyper_ideal(mesh, xm, m, /*energy=*/false).gradient;
+        auto Gm = evaluate_hyper_ideal(mesh, xm, m, /*energy=*/false, /*grad=*/true, clamp).gradient;
 
         xp[sj] = xm[sj] = x[sj];   // restore
 
@@ -101,9 +102,10 @@ inline Eigen::SparseMatrix<double> hyper_ideal_hessian_sym(
     ConformalMesh&             mesh,
     const std::vector<double>& x,
     const HyperIdealMaps&      m,
-    double                     eps = 1e-5)
+    double                     eps   = 1e-5,
+    HyperIdealScaleClamp       clamp = HyperIdealScaleClamp::HardJava)
 {
-    auto H = hyper_ideal_hessian(mesh, x, m, eps);
+    auto H = hyper_ideal_hessian(mesh, x, m, eps, clamp);
     Eigen::SparseMatrix<double> Ht = H.transpose();
     return (H + Ht) * 0.5;
 }
@@ -142,7 +144,8 @@ inline Eigen::SparseMatrix<double> hyper_ideal_hessian_block_fd(
     ConformalMesh&             mesh,
     const std::vector<double>& x,
     const HyperIdealMaps&      m,
-    double                     eps = 1e-5)
+    double                     eps   = 1e-5,
+    HyperIdealScaleClamp       clamp = HyperIdealScaleClamp::HardJava)
 {
     const int n = hyper_ideal_dimension(mesh, m);
     std::vector<Eigen::Triplet<double>> trips;
@@ -187,9 +190,9 @@ inline Eigen::SparseMatrix<double> hyper_ideal_hessian_block_fd(
             vm[j] -= eps;
 
             auto Op = face_angles_from_local_dofs(
-                vp[0], vp[1], vp[2], vp[3], vp[4], vp[5], v1b, v2b, v3b);
+                vp[0], vp[1], vp[2], vp[3], vp[4], vp[5], v1b, v2b, v3b, clamp);
             auto Om = face_angles_from_local_dofs(
-                vm[0], vm[1], vm[2], vm[3], vm[4], vm[5], v1b, v2b, v3b);
+                vm[0], vm[1], vm[2], vm[3], vm[4], vm[5], v1b, v2b, v3b, clamp);
 
             const double Gp[6] = {
                 Op.beta1, Op.beta2, Op.beta3,
@@ -221,9 +224,10 @@ inline Eigen::SparseMatrix<double> hyper_ideal_hessian_block_fd_sym(
     ConformalMesh&             mesh,
     const std::vector<double>& x,
     const HyperIdealMaps&      m,
-    double                     eps = 1e-5)
+    double                     eps   = 1e-5,
+    HyperIdealScaleClamp       clamp = HyperIdealScaleClamp::HardJava)
 {
-    auto H = hyper_ideal_hessian_block_fd(mesh, x, m, eps);
+    auto H = hyper_ideal_hessian_block_fd(mesh, x, m, eps, clamp);
     Eigen::SparseMatrix<double> Ht = H.transpose();
     return (H + Ht) * 0.5;
 }

code/include/newton_solver.hpp

@@ -400,6 +400,12 @@ inline NewtonResult newton_spherical(
 /// \param max_iter  Maximum Newton iterations.  Default: 200.
 /// \param hess_eps  Finite-difference step for Hessian approximation.  Default: 1e-5.
 ///                  (Phase 9b will replace this with an analytic Hessian.)
+/// \param clamp     Vertex-scale floor mode (N3 audit).  Default `HardJava`
+///                  reproduces the Java oracle's hard `b<0 → floor` snap
+///                  (C⁰, parity-faithful).  `SmoothBarrier` uses the C¹
+///                  softplus floor — smoother near the feasibility boundary
+///                  but deviates from the Java golden values.  See
+///                  `HyperIdealScaleClamp` in hyper_ideal_functional.hpp.
 /// \return          NewtonResult{x*, iterations, grad_inf_norm, converged}.
 ///
 /// \see Springborn (2020), Theorem 1.3 for the strict convexity proof.
@@ -410,7 +416,8 @@ inline NewtonResult newton_hyper_ideal(
     const HyperIdealMaps&      m,
     double                     tol      = 1e-8,
     int                        max_iter = 200,
-    double                     hess_eps = 1e-5)
+    double                     hess_eps = 1e-5,
+    HyperIdealScaleClamp       clamp    = HyperIdealScaleClamp::HardJava)
 {
     std::vector<double> x = x0;
     const int n = static_cast<int>(x.size());
@@ -422,7 +429,7 @@ inline NewtonResult newton_hyper_ideal(
 
     for (int iter = 0; iter < max_iter; ++iter) {
         // ── Gradient ──────────────────────────────────────────────────────────
-        auto G_std = evaluate_hyper_ideal(mesh, x, m, /*energy=*/false).gradient;
+        auto G_std = evaluate_hyper_ideal(mesh, x, m, /*energy=*/false, /*grad=*/true, clamp).gradient;
         Eigen::Map<const Eigen::VectorXd> G(G_std.data(), n);
 
         double inf_norm = G.cwiseAbs().maxCoeff();
@@ -435,7 +442,7 @@ inline NewtonResult newton_hyper_ideal(
         }
 
         // ── Hessian (block-FD, ~33–1166× faster than full-FD) + solve H·Δx = −G
-        auto H  = hyper_ideal_hessian_block_fd_sym(mesh, x, m, hess_eps);
+        auto H  = hyper_ideal_hessian_block_fd_sym(mesh, x, m, hess_eps, clamp);
         bool ok = false;
         Eigen::VectorXd dx = detail::solve_with_fallback(H, -G, ok);
         if (!ok) break;
@@ -446,14 +453,14 @@ inline NewtonResult newton_hyper_ideal(
         bool improved = true;
         x = detail::line_search(x, dx, d_sd, norm0,
             [&](const std::vector<double>& xnew) {
-                return evaluate_hyper_ideal(mesh, xnew, m, false).gradient;
+                return evaluate_hyper_ideal(mesh, xnew, m, false, true, clamp).gradient;
             }, &improved);
         if (!improved) break;
 
         res.iterations = iter + 1;
     }
 
-    auto G_final = evaluate_hyper_ideal(mesh, x, m, false).gradient;
+    auto G_final = evaluate_hyper_ideal(mesh, x, m, false, true, clamp).gradient;
     double inf_final = 0.0;
     for (double v : G_final) inf_final = std::max(inf_final, std::abs(v));
     res.grad_inf_norm = inf_final;

code/tests/cgal/test_hyper_ideal_hessian.cpp

@@ -338,3 +338,81 @@ TEST(HyperIdealHessian, BlockFD_FasterThanFullFD)
     EXPECT_GE(ms_full, 3 * ms_block)
         << "Block-FD should be at least 3× faster than full-FD on this mesh";
 }
+
+// ════════════════════════════════════════════════════════════════════════════
+// N3 — Scale-floor clamp modes (HardJava vs SmoothBarrier)
+//
+// The vertex-scale floor can be applied two ways (HyperIdealScaleClamp):
+//   • HardJava (default): b<0 → floor.  Faithful to Java, but only C⁰ — and
+//     in fact value-discontinuous at b=0 (jumps from `floor` to 0).
+//   • SmoothBarrier: floor + softplus_β(b−floor).  C¹ everywhere, ≈ b away
+//     from the floor, and keeps b ≥ floor > 0 smoothly.
+// These tests pin the contract of both modes and the default-preservation.
+// ════════════════════════════════════════════════════════════════════════════
+
+TEST(HyperIdealClamp, SmoothBarrierIsContinuousAndC1AcrossZero)
+{
+    const auto Hard   = HyperIdealScaleClamp::HardJava;
+    const auto Smooth = HyperIdealScaleClamp::SmoothBarrier;
+    auto f = [](double b, HyperIdealScaleClamp mode) {
+        return clamp_hyper_ideal_scale(b, /*variable=*/true, mode);
+    };
+
+    const double eps = 1e-6;
+
+    // HardJava jumps in VALUE at b=0 (floor on the left, 0 on the right).
+    EXPECT_GT(std::abs(f(-eps, Hard) - f(+eps, Hard)), 0.5 * HYPER_IDEAL_SCALE_FLOOR);
+
+    // SmoothBarrier is continuous in value across b=0 …
+    EXPECT_NEAR(f(-eps, Smooth), f(+eps, Smooth), 1e-4);
+
+    // … and C¹: the one-sided slopes across b=0 agree (the HardJava slopes,
+    // 0 on the left and 1 on the right, would not).
+    auto slope = [&](double b0, HyperIdealScaleClamp mode) {
+        return (f(b0 + eps, mode) - f(b0 - eps, mode)) / (2.0 * eps);
+    };
+    const double sL = slope(-1e-3, Smooth);
+    const double sR = slope(+1e-3, Smooth);
+    EXPECT_NEAR(sL, sR, 5e-2) << "SmoothBarrier derivative should be continuous";
+}
+
+TEST(HyperIdealClamp, SmoothBarrierStaysAboveFloorAndApproachesIdentity)
+{
+    const auto Smooth = HyperIdealScaleClamp::SmoothBarrier;
+    auto f = [&](double b) { return clamp_hyper_ideal_scale(b, true, Smooth); };
+
+    // At or above the floor for any input (mathematically > floor; for very
+    // negative b the softplus underflows to 0 in double, giving exactly floor).
+    EXPECT_GE(f(-1e6), HYPER_IDEAL_SCALE_FLOOR);
+    EXPECT_GE(f(-1.0), HYPER_IDEAL_SCALE_FLOOR);
+    EXPECT_GT(f(0.0),  0.0);
+    EXPECT_LT(f(-1e6) - HYPER_IDEAL_SCALE_FLOOR, 1e-9);   // converges down to floor
+
+    // ≈ identity well above the floor (the normal operating regime).
+    EXPECT_NEAR(f(1.0),  1.0,  1e-9);
+    EXPECT_NEAR(f(5.0),  5.0,  1e-12);
+
+    // Pinned DOFs are never clamped, regardless of mode.
+    EXPECT_EQ(clamp_hyper_ideal_scale(-3.0, /*variable=*/false, Smooth), -3.0);
+}
+
+// Away from the feasibility boundary (all b = 1 ≫ floor), the two modes must
+// produce the same Hessian — SmoothBarrier only differs near b ≈ floor, so the
+// default-mode parity results are not perturbed in the normal regime.
+TEST(HyperIdealClamp, ModesAgreeAwayFromBoundary)
+{
+    auto mesh = make_tetrahedron();
+    auto m    = setup_hyper_ideal_maps(mesh);
+    const int n = assign_all_dof_indices(mesh, m);
+    auto x = natural_x(mesh, m);   // b = 1, a = 0.5 — all well above the floor
+
+    auto H_hard = hyper_ideal_hessian_block_fd_sym(
+        mesh, x, m, 1e-5, HyperIdealScaleClamp::HardJava);
+    auto H_soft = hyper_ideal_hessian_block_fd_sym(
+        mesh, x, m, 1e-5, HyperIdealScaleClamp::SmoothBarrier);
+
+    Eigen::MatrixXd Dh(H_hard), Ds(H_soft);
+    EXPECT_LT((Dh - Ds).cwiseAbs().maxCoeff(), 1e-9)
+        << "clamp modes must agree when every scale is well above the floor";
+    (void)n;
+}

code/tests/cgal/test_lawson_hyperideal.cpp

@@ -194,6 +194,54 @@ TEST(LawsonHyperIdeal, ConvergenceGoldenVector_JavaXVal)
     }
 }
 
+// ════════════════════════════════════════════════════════════════════════════
+// N3 — SmoothBarrier clamp mode converges to the same golden solution
+//
+// The C¹ smooth-barrier scale floor is an opt-in alternative to the Java hard
+// clamp.  On this well-posed problem the scales stay well above the floor
+// throughout the solve, so the barrier is ≈ identity and the converged vector
+// must match the same golden classes as the HardJava run above — demonstrating
+// the smooth mode is a drop-in that does not move the solution when the clamp
+// region is never entered.
+// ════════════════════════════════════════════════════════════════════════════
+TEST(LawsonHyperIdeal, SmoothBarrierConvergesToGoldenVector)
+{
+    std::set<Edge_index> original;
+    ConformalMesh m = make_lawson_square_tiled(&original);
+
+    HyperIdealMaps maps = setup_hyper_ideal_maps(m);
+    const int n = assign_all_dof_indices(m, maps);
+    ASSERT_EQ(n, 4 + 18);
+
+    for (auto e : m.edges())
+        if (original.count(e)) maps.theta_e[e] = PI / 2.0;
+
+    std::vector<double> x0(static_cast<std::size_t>(n), 1.0);
+    auto res = newton_hyper_ideal(m, x0, maps, /*tol=*/1e-10, /*max_iter=*/200,
+                                  /*hess_eps=*/1e-5,
+                                  HyperIdealScaleClamp::SmoothBarrier);
+    ASSERT_TRUE(res.converged)
+        << "SmoothBarrier HyperIdeal Newton did not converge; ||G||="
+        << res.grad_inf_norm;
+
+    constexpr double b_gold      = 1.1462158341786262;
+    constexpr double a_orig_gold = 1.7627471737467797;
+    constexpr double a_aux_gold  = 2.633915794495759;
+    const double tol = 1e-5;
+
+    for (auto v : m.vertices()) {
+        const int iv = maps.v_idx[v];
+        ASSERT_GE(iv, 0);
+        EXPECT_NEAR(res.x[static_cast<std::size_t>(iv)], b_gold, tol);
+    }
+    for (auto e : m.edges()) {
+        const int ie = maps.e_idx[e];
+        ASSERT_GE(ie, 0);
+        const double gold = original.count(e) ? a_orig_gold : a_aux_gold;
+        EXPECT_NEAR(res.x[static_cast<std::size_t>(ie)], gold, tol);
+    }
+}
+
 // ════════════════════════════════════════════════════════════════════════════
 // Branch-points variant — Java HyperIdealConvergenceTest...WithBranchPoints
 //

doc/math/geometry-modes.md

@@ -80,8 +80,27 @@ lᵢⱼ(ζ)                             — edge length from ζ value
 βᵢ(l₁, l₂, l₃)                    — vertex angle sum contribution
 ```
 
-**Hessian:** currently a symmetric finite-difference approximation (see `hyper_ideal_hessian.hpp`).
+**Hessian:** a symmetric finite-difference approximation (see `hyper_ideal_hessian.hpp`).
-The analytic Hessian through the `(bᵢ, aₑ) → lᵢⱼ → ζ → αᵢⱼ/βᵢ` chain is Phase 9b.
+The solver uses the **per-face block-FD** variant by default (`O(F)` face evaluations,
+~33–1166× faster than full-FD); the full-FD baseline is retained for cross-validation.
+The analytic Hessian through the `(bᵢ, aₑ) → lᵢⱼ → ζ → αᵢⱼ/βᵢ` chain is Phase 9b-analytic.
+
+**Scale-floor clamp mode (`HyperIdealScaleClamp`).** The vertex scale `bᵢ` must stay
+positive for the geometry to be valid. When a Newton iterate drives `bᵢ` toward or below
+zero, one of two floors is applied — selectable per solver call (last argument of
+`newton_hyper_ideal`, default `HardJava`):
+
+| Mode | Behaviour | When to use |
+|---|---|---|
+| `HardJava` *(default)* | `bᵢ < 0 → HYPER_IDEAL_SCALE_FLOOR` (= 0.01). Bit-for-bit faithful to `HyperIdealFunctional.java`, so the Java golden-oracle parity tests hold **exactly**. It is only C⁰ (in fact value-discontinuous at `bᵢ = 0`), so a step crossing the feasibility boundary sees a kink that can stall Newton. | Default — preserves parity; correct whenever scales stay positive (the normal case). |
+| `SmoothBarrier` | `bᵢ ↦ floor + softplusβ(bᵢ − floor)` with `β = HYPER_IDEAL_SCALE_SHARPNESS` (= 100). C¹ everywhere, ≈ `bᵢ` to machine precision once `β·(bᵢ−floor) ≳ 35`, and decays smoothly to `floor` as `bᵢ → −∞`. Removes the N3 kink, but perturbs values near the boundary and therefore **deviates from the Java oracle** there. | Opt in when a solve is expected to cross the feasibility boundary and convergence robustness matters more than strict Java parity. |
+
+Both modes share `HYPER_IDEAL_SCALE_FLOOR`; the edge clamp (`aₑ < 0 → 0`) is unchanged.
+Away from the floor (the normal `bᵢ ≫ 0.01` regime) the two modes are numerically
+identical — the `LawsonHyperIdeal.SmoothBarrierConvergesToGoldenVector` test confirms the
+smooth mode reaches the same golden solution as `HardJava`. Implemented in
+`clamp_hyper_ideal_scale` (`hyper_ideal_functional.hpp`); origin: numerical-stability
+audit finding **N3**.
 
 **Holonomy:** Möbius isometries Tᵢ ∈ SU(1,1) (orientation-preserving isometries of the Poincaré disk).
 `MobiusMap` in `layout.hpp`: T(z) = (az+b)/(cz+d).