diff --git a/examples/Solovev_kinetic_calculated_example/gpec.toml b/examples/Solovev_kinetic_calculated_example/gpec.toml
index 4a46bdb4..15ce5854 100644
--- a/examples/Solovev_kinetic_calculated_example/gpec.toml
+++ b/examples/Solovev_kinetic_calculated_example/gpec.toml
@@ -1,6 +1,10 @@
+# Solovev analytical equilibrium — n=1 ideal stability with calculated kinetic matrices.
+# Exercises the real KineticForces (NTV) physics path at full strength; kinetic regression
+# fixture (et[1]≈1.894-1.525i). The equilibrium is generated analytically from the embedded
+# [SOL_INPUT] section, with n·T tied to the Solovev P₀≈4.27e4 Pa via the local kinetic.dat.
+
 [Equilibrium]
 eq_type = "sol"                         # Type of the input 2D equilibrium file
-eq_filename = "sol.toml"                # Path to equilibrium file
 jac_type = "pest"                       # Coordinate system (hamada, pest, boozer, equal_arc, park, custom)
 grid_type = "ldp"                       # Radial grid packing type
 psilow = 1e-4                           # Lower limit of normalized flux coordinate
@@ -45,3 +49,16 @@ singfac_min = 1e-4            # Fractional distance from rational q at which ide
 ucrit = 1e3                   # Maximum fraction of solutions allowed before re-normalized
 write_outputs_to_HDF5 = true
 verbose = false
+
+# Solovev analytic equilibrium parameters (eq_type = "sol"); see SolovevConfig in src/Equilibrium.
+[SOL_INPUT]
+mr = 128       # Number of radial grid zones
+mz = 128       # Number of axial grid zones
+ma = 128       # Number of flux grid zones
+e = 1.6        # Elongation
+a = 0.33       # Minor radius
+r0 = 1.0       # Major radius
+q0 = 1.9       # Safety factor at the magnetic axis (O-point)
+p0fac = 1      # Scale on-axis pressure (β changes; Φ, q fixed)
+b0fac = 1      # Scale toroidal field at constant β (Bt changes; shape, β fixed)
+f0fac = 1      # Scale toroidal field at constant pressure (β, q change; Φ, p, Bp fixed)
diff --git a/examples/Solovev_kinetic_calculated_example/sol.toml b/examples/Solovev_kinetic_calculated_example/sol.toml
deleted file mode 100644
index e54fdeac..00000000
--- a/examples/Solovev_kinetic_calculated_example/sol.toml
+++ /dev/null
@@ -1,11 +0,0 @@
-[SOL_INPUT]
-mr = 128      # number of radial grid zones
-mz = 128      # number of axial grid zones
-ma = 128      # number of flux grid zones
-e = 1.6       # elongation
-a = 0.33      # minor radius
-r0 = 1.0      # major radius
-q0 = 1.9      # safety factor at the o-point
-p0fac=1       # scale on-axis pressure (P-> P+P0*p0fac. beta changes. Phi,q constant)
-b0fac=1       # scale toroidal field at constant beta (s*Phi,s*f,s^2*P. bt changes. Shape,beta constant)
-f0fac=1       # scale toroidal field at constant pressure (s*f. beta,q changes. Phi,p,bp constant)
diff --git a/src/KineticForces/BounceAveraging.jl b/src/KineticForces/BounceAveraging.jl
index 3a75b003..e0153fcb 100644
--- a/src/KineticForces/BounceAveraging.jl
+++ b/src/KineticForces/BounceAveraging.jl
@@ -204,6 +204,9 @@ function compute_bounce_data(
     mpert = length(mfac)
     do_matrices = !isnothing(smat)
 
+    # Per-surface scratch, reused across all λ.
+    scr = BounceScratch(ntheta, mpert)
+
     # Trapped-passing boundary and λ range
     lmdatpb = bo / bmax
     lmdamax = bo / bmin
@@ -238,13 +241,13 @@ function compute_bounce_data(
         # Find bounce points and build θ sub-grid
         _, _, tdt_pts, tdt_wts = _find_bounce_points_and_grid(
             lmda, bo, sigma, tspl, ibmax, theta_bmax,
-            lmdatpb, lmdamax, psi, ntheta)
+            lmdatpb, lmdamax, psi, ntheta, scr)
 
         # Bounce integrals over θ (Fortran lines 674-735)
         wbbar, wdbar, dJdJ_val, wmats_lmda = _bounce_integrate(
             tdt_pts, tdt_wts, lmda, lnq, sigma, n, q, bo,
             tspl, chi1, ro, mfac, dbob_m_f, divx_m_f, divxfac, wdfac,
-            do_matrices, mpert, smat, tmat, xmat, ymat, zmat)
+            do_matrices, mpert, smat, tmat, xmat, ymat, zmat, scr)
 
         # Physical frequencies (Fortran lines 744-745)
         wb_arr[ilmda] = wbbar * bhat
@@ -263,6 +266,60 @@ function compute_bounce_data(
 end
 
 
+"""
+    BounceScratch(ntheta, mpert; nfine=256)
+
+Per-surface scratch buffers for the bounce-averaging inner loops. Allocated once in
+`compute_bounce_data` and reused across every λ (each λ previously reallocated these
+θ- and mode-sized arrays). Sizes are fixed for a flux surface (`ntheta` sub-grid points,
+`mpert` Fourier modes), so a single set of buffers serves the whole λ sweep. Buffers that
+the callers zero-initialize are `fill!`-reset per λ, preserving bit-for-bit results.
+
+## Fields
+- `cum_wb_arr::Vector{Float64}`: length `ntheta` — cumulative bounce action
+- `jvtheta::Vector{ComplexF64}`: length `ntheta` — action integrand
+- `wmu_mt::Matrix{ComplexF64}`: `mpert × ntheta` — W_μ per θ
+- `wen_mt::Matrix{ComplexF64}`: `mpert × ntheta` — W_E per θ
+- `expm::Vector{ComplexF64}`: length `mpert` — Fourier basis at a θ
+- `pl::Vector{ComplexF64}`: length `ntheta` — bounce phase factor
+- `wmu_ba::Vector{ComplexF64}`: length `mpert` — bounce-averaged W_μ
+- `wen_ba::Vector{ComplexF64}`: length `mpert` — bounce-averaged W_E
+- `wmats_lmda::Vector{ComplexF64}`: length `nqty_matrix(mpert)` — packed W outer products
+- `tspl_f::Vector{Float64}`: length 5 — in-place tspl(θ) evaluation
+- `vpar_fine::Vector{Float64}`: length `nfine+1` — v_par(θ) for bounce-point search;
+  `_find_bounce_points_and_grid` derives its fine-grid size from this buffer's length
+"""
+struct BounceScratch
+    cum_wb_arr::Vector{Float64}
+    jvtheta::Vector{ComplexF64}
+    wmu_mt::Matrix{ComplexF64}
+    wen_mt::Matrix{ComplexF64}
+    expm::Vector{ComplexF64}
+    pl::Vector{ComplexF64}
+    wmu_ba::Vector{ComplexF64}
+    wen_ba::Vector{ComplexF64}
+    wmats_lmda::Vector{ComplexF64}
+    tspl_f::Vector{Float64}
+    vpar_fine::Vector{Float64}
+end
+
+function BounceScratch(ntheta::Int, mpert::Int; nfine::Int=256)
+    return BounceScratch(
+        Vector{Float64}(undef, ntheta),
+        Vector{ComplexF64}(undef, ntheta),
+        Matrix{ComplexF64}(undef, mpert, ntheta),
+        Matrix{ComplexF64}(undef, mpert, ntheta),
+        Vector{ComplexF64}(undef, mpert),
+        Vector{ComplexF64}(undef, ntheta),
+        Vector{ComplexF64}(undef, mpert),
+        Vector{ComplexF64}(undef, mpert),
+        Vector{ComplexF64}(undef, nqty_matrix(mpert)),
+        Vector{Float64}(undef, 5),
+        Vector{Float64}(undef, nfine + 1),
+    )
+end
+
+
 # ============================================================================
 # Internal helpers
 # ============================================================================
@@ -309,21 +366,25 @@ function _find_bounce_points_and_grid(
     lmda::Float64, bo::Float64, sigma::Int,
     tspl, ::Int, theta_bmax::Float64,
     ::Float64, ::Float64, psi::Float64,
-    ntheta::Int
+    ntheta::Int, scr::BounceScratch
 )
     if sigma == 0  # trapped
-        # Build v_par(θ) = 1 - (λ/bo)*B(θ) and find roots
-        # Use a dense θ grid to find zero crossings
-        nfine = 256
+        # Build v_par(θ) = 1 - (λ/bo)*B(θ) and find roots on a dense θ grid.
+        # Grid size derives from the scratch buffer so it stays in sync with BounceScratch's nfine.
+        nfine = length(scr.vpar_fine) - 1
         theta_fine = range(0.0, 1.0, length=nfine+1)
-        vpar_fine = [1.0 - (lmda / bo) * tspl(mod(θ, 1.0))[1] for θ in theta_fine]
+        vpar_fine = scr.vpar_fine
+        @inbounds for i in 1:(nfine+1)
+            tspl(scr.tspl_f, mod(theta_fine[i], 1.0))
+            vpar_fine[i] = 1.0 - (lmda / bo) * scr.tspl_f[1]
+        end
 
         # Find zero crossings
         bpts = Float64[]
         for i in 1:nfine
             if vpar_fine[i] * vpar_fine[i+1] < 0
                 # Bisect for better accuracy
-                θ_root = _bisect_vpar(tspl, lmda, bo, theta_fine[i], theta_fine[i+1])
+                θ_root = _bisect_vpar(tspl, scr.tspl_f, lmda, bo, theta_fine[i], theta_fine[i+1])
                 push!(bpts, θ_root)
             end
         end
@@ -339,7 +400,7 @@ function _find_bounce_points_and_grid(
             t2 = bpts[1] + 1.0
         else
             # Find deepest potential well (Fortran lines 616-639)
-            t1, t2 = _find_deepest_well(bpts, tspl, lmda, bo)
+            t1, t2 = _find_deepest_well(bpts, tspl, scr.tspl_f, lmda, bo)
         end
 
         # Power-law grid refined near bounce points
@@ -381,11 +442,13 @@ this routine does not check, it just halves toward the sign change.
 # Returns
 - `θ::Float64`: the converged (or capped) bounce-point angle.
 """
-function _bisect_vpar(tspl, lmda::Float64, bo::Float64, θa::Float64, θb::Float64; tol=1e-12, maxiter=50)
-    va = 1.0 - (lmda / bo) * tspl(mod(θa, 1.0))[1]
+function _bisect_vpar(tspl, tspl_f::Vector{Float64}, lmda::Float64, bo::Float64, θa::Float64, θb::Float64; tol=1e-12, maxiter=50)
+    tspl(tspl_f, mod(θa, 1.0))
+    va = 1.0 - (lmda / bo) * tspl_f[1]
     for _ in 1:maxiter
         θm = 0.5 * (θa + θb)
-        vm = 1.0 - (lmda / bo) * tspl(mod(θm, 1.0))[1]
+        tspl(tspl_f, mod(θm, 1.0))
+        vm = 1.0 - (lmda / bo) * tspl_f[1]
         if abs(vm) < tol || (θb - θa) < tol
             return θm
         end
@@ -404,7 +467,7 @@ end
 Find deepest potential well among bounce point pairs.
 Ports Fortran lines 616-639.
 """
-function _find_deepest_well(bpts::Vector{Float64}, tspl, lmda::Float64, bo::Float64)
+function _find_deepest_well(bpts::Vector{Float64}, tspl, tspl_f::Vector{Float64}, lmda::Float64, bo::Float64)
     nbpts = length(bpts)
     best_vpar = 0.0
     best_t1 = 0.0
@@ -418,7 +481,8 @@ function _find_deepest_well(bpts::Vector{Float64}, tspl, lmda::Float64, bo::Floa
         else
             θmid = 0.5 * (bpts[i] + bpts[j])
         end
-        vpar_mid = 1.0 - (lmda / bo) * tspl(mod(θmid, 1.0))[1]
+        tspl(tspl_f, mod(θmid, 1.0))
+        vpar_mid = 1.0 - (lmda / bo) * tspl_f[1]
         if vpar_mid > best_vpar
             best_t1 = bpts[i]
             best_t2 = bpts[j]
@@ -451,7 +515,7 @@ function _bounce_integrate(
     mfac::Vector{Int}, dbob_m_f::Vector{ComplexF64}, divx_m_f::Vector{ComplexF64},
     divxfac::Float64, wdfac::Float64,
     do_matrices::Bool, mpert::Int,
-    smat, tmat, xmat, ymat, zmat
+    smat, tmat, xmat, ymat, zmat, scr::BounceScratch
 )
     ntheta = length(tdt_pts)
     theta0 = tdt_pts[1]
@@ -460,23 +524,27 @@ function _bounce_integrate(
     cum_wb = 0.0
     cum_wd = 0.0
 
-    # θ-scratch array allocated fresh each call. Pool-based reuse was tried
-    # (AdaptiveArrayPools) but showed no speedup and introduced severe slowdowns
-    # at 2+ threads; plain allocations match Fortran baseline behavior.
-    cum_wb_arr = zeros(Float64, ntheta)
+    # θ-scratch arrays reused across λ from the per-surface BounceScratch. Zero-reset
+    # the ones the loop populates only partially (i in 2:ntheta-1, with continue/break),
+    # so unwritten entries stay 0 exactly as the previous `zeros(...)` did.
+    cum_wb_arr = scr.cum_wb_arr
+    fill!(cum_wb_arr, 0.0)
 
     # Action integrand
-    jvtheta = zeros(ComplexF64, ntheta)
+    jvtheta = scr.jvtheta
+    fill!(jvtheta, ComplexF64(0.0))
 
     # W vectors for matrix path
-    wmu_mt = do_matrices ? zeros(ComplexF64, mpert, ntheta) : nothing
-    wen_mt = do_matrices ? zeros(ComplexF64, mpert, ntheta) : nothing
+    wmu_mt = scr.wmu_mt
+    wen_mt = scr.wen_mt
+    if do_matrices
+        fill!(wmu_mt, ComplexF64(0.0))
+        fill!(wen_mt, ComplexF64(0.0))
+    end
 
-    # Pre-allocated scratch for hot-loop tspl evaluation + Fourier-basis buffer
-    # (avoids Vector{Float64}(5) + Vector{ComplexF64}(mpert) allocation per θ
-    # sub-grid point — previously ~256 allocs × 126 inner iters per call).
-    tspl_f = Vector{Float64}(undef, 5)
-    expm = Vector{ComplexF64}(undef, mpert)
+    # Scratch for hot-loop tspl evaluation + Fourier-basis buffer (fully written per use).
+    tspl_f = scr.tspl_f
+    expm = scr.expm
 
     for i in 2:ntheta-1  # Edge weights are 0 from powspace
         θ = tdt_pts[i]
@@ -579,7 +647,7 @@ function _bounce_integrate(
     one_minus_sigma = 1 - sigma
     bj_integral = ComplexF64(0.0)
     if do_matrices
-        pl = Vector{ComplexF64}(undef, ntheta)
+        pl = scr.pl
         @inbounds for i in 1:ntheta
             pl[i] = cis(-twopi * lnq * cum_wb_arr[i] * nrm / pl_denom)
         end
@@ -606,8 +674,10 @@ function _bounce_integrate(
         # Bounce-average W_μ and W_E vectors (Fortran lines 762-767).
         # Trapezoidal quadrature: boundary samples weighted by 0.5 (wmu_mt and wen_mt
         # are zero at i=1 and i=ntheta from the 2:ntheta-1 population loop above).
-        wmu_ba = zeros(ComplexF64, mpert)
-        wen_ba = zeros(ComplexF64, mpert)
+        wmu_ba = scr.wmu_ba
+        wen_ba = scr.wen_ba
+        fill!(wmu_ba, ComplexF64(0.0))
+        fill!(wen_ba, ComplexF64(0.0))
         @inbounds for i in 1:ntheta
             w = (i == 1 || i == ntheta) ? 0.5 : 1.0
             factor = w * (pl[i] + one_minus_sigma / (pl[i] + SINGULAR_EPS))
@@ -636,7 +706,9 @@ function _bounce_integrate(
         # Scale by wbbar/ro² (Fortran line 789)
         scale = wbbar / ro^2
         Mu = (mpert * (mpert + 1)) ÷ 2
-        wmats_lmda = Vector{ComplexF64}(undef, nqty_matrix(mpert))
+        wmats_lmda = scr.wmats_lmda
+        # Fill with NaN to catch uninitialized entries
+        fill!(wmats_lmda, ComplexF64(NaN, NaN))
 
         # A (Hermitian): upper triangle of W_Z†W_Z, rank-1 → conj(wz[i])·wz[j].
         off = 0
diff --git a/src/KineticForces/EnergyIntegration.jl b/src/KineticForces/EnergyIntegration.jl
index 7c441582..36cc1ece 100644
--- a/src/KineticForces/EnergyIntegration.jl
+++ b/src/KineticForces/EnergyIntegration.jl
@@ -208,9 +208,34 @@ The collisionless case (ν ≡ 0, real-axis pole) is the exact ν→0⁺ limit o
 formula; its causal branch is carried by the signed zero of `−pole_offset` (see the
 add-back below), and its on-axis `0/0` window by the analytic regular-part limit.
 """
+# Real x-space resonant integrand with pole subtractions
+# Explicit function keeps memory allocation out of the QuadGK inner loop.
+@inline function _resonant_integrand(x::Float64, p::EnergyParams,
+        residues::Vector{ComplexF64}, x_poles::Vector{ComplexF64}, npole::Int,
+        leff::Float64, wb::Float64, n::Int, wd::Float64)::ComplexF64
+    val = _energy_integrand_real(x, p)
+    @inbounds for k in 1:npole
+        val -= residues[k] / (x - x_poles[k])
+    end
+    # At a ν=0 pole Ω rounds to 0 and the physical term is 0/0; substitute the nearest
+    # pole's analytic regular-part limit, keeping the other (finite) pole subtractions.
+    if !isfinite(val)
+        k = 1
+        @inbounds for j in 2:npole
+            abs(x - real(x_poles[j])) < abs(x - real(x_poles[k])) && (k = j)
+        end
+        val = _real_pole_regular_part(real(x_poles[k]), p, leff, wb, n, wd)
+        @inbounds for j in 1:npole
+            j == k && continue
+            val -= residues[j] / (x - x_poles[j])
+        end
+    end
+    return val
+end
+
 function _integrate_energy_resonant(p::EnergyParams, leff::Float64, wb::Float64,
                                         n::Int, we::Float64, wd::Float64,
-                                        atol::Float64, rtol::Float64)::ComplexF64
+                                        atol::Float64, rtol::Float64, segbuf=nothing)::ComplexF64
     x_res_list = find_resonance_energies(leff, wb, n, we, wd)   # ≤ 2 roots of a quadratic in √x
     x_poles = ComplexF64[]
     residues = ComplexF64[]
@@ -240,29 +265,10 @@ function _integrate_energy_resonant(p::EnergyParams, leff::Float64, wb::Float64,
     npole = length(residues)
 
     # Physical x-space integrand N(x)·exp(-x)/(i·Ω) minus the subtracted pole parts.
-    integrand = x -> begin
-        val = _energy_integrand_real(x, p)
-        @inbounds for k in 1:npole
-            val -= residues[k] / (x - x_poles[k])
-        end
-        # At a ν=0 pole Ω rounds to 0 and the physical term is 0/0; substitute the nearest
-        # pole's analytic regular-part limit, keeping the other (finite) pole subtractions.
-        if !isfinite(val)
-            k = 1
-            @inbounds for j in 2:npole
-                abs(x - real(x_poles[j])) < abs(x - real(x_poles[k])) && (k = j)
-            end
-            val = _real_pole_regular_part(real(x_poles[k]), p, leff, wb, n, wd)
-            @inbounds for j in 1:npole
-                j == k && continue
-                val -= residues[j] / (x - x_poles[j])
-            end
-        end
-        return val
-    end
+    integrand = x -> _resonant_integrand(x, p, residues, x_poles, npole, leff, wb, n, wd)
 
     if npole == 0
-        val, _ = quadgk(integrand, 0.0, X_ENERGY_MAX; atol=atol, rtol=rtol)
+        val, _ = quadgk(integrand, 0.0, X_ENERGY_MAX; atol=atol, rtol=rtol, segbuf=segbuf)
         return val
     end
 
@@ -271,9 +277,9 @@ function _integrate_energy_resonant(p::EnergyParams, leff::Float64, wb::Float64,
     # positional args — a `breaks...` splat of a runtime-length Vector is type-unstable.
     breaks = unique(sort(real.(x_poles)))
     smooth_val, _ = if length(breaks) == 1
-        quadgk(integrand, 0.0, breaks[1], X_ENERGY_MAX; atol=atol, rtol=rtol)
+        quadgk(integrand, 0.0, breaks[1], X_ENERGY_MAX; atol=atol, rtol=rtol, segbuf=segbuf)
     else
-        quadgk(integrand, 0.0, breaks[1], breaks[2], X_ENERGY_MAX; atol=atol, rtol=rtol)
+        quadgk(integrand, 0.0, breaks[1], breaks[2], X_ENERGY_MAX; atol=atol, rtol=rtol, segbuf=segbuf)
     end
     return smooth_val + pole_contribution
 end
@@ -309,7 +315,7 @@ function integrate_energy(wn::Float64, wt::Float64, we::Float64, wd::Float64,
                               n::Int, psi::Float64, lambda::Float64, method::String;
                               nutype::String="harmonic", f0type::String="maxwellian",
                               nufac::Float64=1.0, ximag::Float64=0.0, qt::Bool=false,
-                              atol::Float64=1e-7, rtol::Float64=1e-5)::ComplexF64
+                              atol::Float64=1e-7, rtol::Float64=1e-5, segbuf=nothing)::ComplexF64
 
     p = EnergyParams(wn, wt, we, wd, wb, nuk, leff, n,
                      nutype, f0type, nufac, ximag, qt)
@@ -317,13 +323,13 @@ function integrate_energy(wn::Float64, wt::Float64, we::Float64, wd::Float64,
     # CGL has no resonance denominator and no pole — integrate the physical
     # x-space integrand directly over the half line (QuadGK maps [0,∞) itself).
     if f0type == "cgl"
-        val, _ = quadgk(x -> _energy_integrand_real(x, p), 0.0, Inf; atol=atol, rtol=rtol)
+        val, _ = quadgk(x -> _energy_integrand_real(x, p), 0.0, Inf; atol=atol, rtol=rtol, segbuf=segbuf)
         return val
     end
 
     # Single resonant path for any collisionality: real x-space PV+residue with a
     # pole x_pole = x_res - i·ν/Ω′ (off-axis for ν>0, real for ν≡0).
-    return _integrate_energy_resonant(p, leff, wb, n, we, wd, atol, rtol)
+    return _integrate_energy_resonant(p, leff, wb, n, we, wd, atol, rtol, segbuf)
 end
 
 """
diff --git a/src/KineticForces/PitchIntegration.jl b/src/KineticForces/PitchIntegration.jl
index a66091c7..fa61c595 100644
--- a/src/KineticForces/PitchIntegration.jl
+++ b/src/KineticForces/PitchIntegration.jl
@@ -22,7 +22,7 @@ Parameters for the GAR pitch-angle integrand.
 Uses a unified fbnce interpolant that returns [ωb, ωd, f₁, f₂, ...] at each λ,
 matching Fortran `lambdaintgrl_lsode`.
 """
-struct PitchGARParams
+struct PitchGARParams{F}
     wn::Float64        # density gradient diamagnetic drift frequency
     wt::Float64        # temperature gradient diamagnetic drift frequency
     we::Float64        # electric precession frequency
@@ -44,9 +44,11 @@ struct PitchGARParams
     rex::Float64       # real part multiplier for resonance operator
     imx::Float64       # imaginary part multiplier for resonance operator
     nqty::Int          # number of flux quantities to integrate
-    fbnce::Any         # CubicSeriesInterpolant: fbnce(λ) → [wb, wd, f₁, ...]
+    fbnce::F           # CubicSeriesInterpolant: fbnce(λ) → [wb, wd, f₁, ...] (typed for stability)
     fbnce_norm::Vector{Float64}  # normalization factors (1/median)
     fbnce_hint::Base.RefValue{Int}  # sticky bracket-search hint for fbnce(λ)
+    fvals::Vector{ComplexF64}    # reusable buffer for the in-place fbnce(fvals, λ) evaluation; must be complex — the wt-path fbnce carries complex op_wmats data
+    esegbuf::Vector{QuadGK.Segment{Float64,ComplexF64,Float64}}  # reusable energy-integral QuadGK segment buffer
 end
 
 
@@ -84,7 +86,9 @@ function integrate_pitch_gar_quadgk(
         wn, wt, we, nuk, bobmax, epsr, q, ell, n, psi, method,
         nutype, f0type, nufac, ximag, qt,
         energy_atol, energy_rtol,
-        rex, imx, nqty, fbnce, fbnce_norm, Ref(1))
+        rex, imx, nqty, fbnce, fbnce_norm, Ref(1),
+        Vector{ComplexF64}(undef, 2 + nqty),
+        QuadGK.alloc_segbuf(Float64, ComplexF64, Float64))
 
     lambda_min = first(fbnce.cache.x)
     lambda_max = last(fbnce.cache.x)
@@ -114,7 +118,8 @@ In-place complex-valued kernel for `quadgk!`. Writes `out[i] = fvals[i+2] * xint
 for i in 1..nqty. QuadGK natively handles ComplexF64.
 """
 function _pitch_gar_kernel_quadgk!(out::Vector{ComplexF64}, lambda, p::PitchGARParams)
-    fvals = p.fbnce(lambda; hint=p.fbnce_hint)
+    p.fbnce(p.fvals, lambda; hint=p.fbnce_hint)
+    fvals = p.fvals
     wb = real(fvals[1])
     wd = real(fvals[2])
 
@@ -127,12 +132,12 @@ function _pitch_gar_kernel_quadgk!(out::Vector{ComplexF64}, lambda, p::PitchGARP
                                         p.ell, leff, p.n, p.psi, lambda, p.method;
                                         nutype=p.nutype, f0type=p.f0type,
                                         nufac=p.nufac, ximag=p.ximag, qt=p.qt,
-                                        atol=p.energy_atol, rtol=p.energy_rtol)
+                                        atol=p.energy_atol, rtol=p.energy_rtol, segbuf=p.esegbuf)
         xint_counter = integrate_energy(p.wn, p.wt, p.we, wd, -wb, nueff,
                                              p.ell, leff, p.n, p.psi, lambda, p.method;
                                              nutype=p.nutype, f0type=p.f0type,
                                              nufac=p.nufac, ximag=p.ximag, qt=p.qt,
-                                             atol=p.energy_atol, rtol=p.energy_rtol)
+                                             atol=p.energy_atol, rtol=p.energy_rtol, segbuf=p.esegbuf)
         xint = xint_co + xint_counter
     else
         xint = integrate_energy(p.wn, p.wt, p.we, wd, wb, nueff,
@@ -182,7 +187,9 @@ function integrate_pitch_gar_quadgk_wt(
         wn, wt, we, nuk, bobmax, epsr, q, ell, n, psi, method,
         nutype, f0type, nufac, ximag, qt,
         energy_atol, energy_rtol,
-        1.0, 1.0, nqty, fbnce, fbnce_norm, Ref(1))
+        1.0, 1.0, nqty, fbnce, fbnce_norm, Ref(1),
+        Vector{ComplexF64}(undef, 2 + nqty),
+        QuadGK.alloc_segbuf(Float64, ComplexF64, Float64))
 
     lambda_min = first(fbnce.cache.x)
     lambda_max = last(fbnce.cache.x)
@@ -211,7 +218,8 @@ Dual-output pitch kernel. Fills a length-`2*nqty` buffer:
 One energy integration per λ; both halves share it.
 """
 function _pitch_gar_kernel_quadgk_wt!(out::Vector{ComplexF64}, lambda, p::PitchGARParams)
-    fvals = p.fbnce(lambda; hint=p.fbnce_hint)
+    p.fbnce(p.fvals, lambda; hint=p.fbnce_hint)
+    fvals = p.fvals
     wb = real(fvals[1])
     wd = real(fvals[2])
 
@@ -224,12 +232,12 @@ function _pitch_gar_kernel_quadgk_wt!(out::Vector{ComplexF64}, lambda, p::PitchG
                                         p.ell, leff, p.n, p.psi, lambda, p.method;
                                         nutype=p.nutype, f0type=p.f0type,
                                         nufac=p.nufac, ximag=p.ximag, qt=p.qt,
-                                        atol=p.energy_atol, rtol=p.energy_rtol)
+                                        atol=p.energy_atol, rtol=p.energy_rtol, segbuf=p.esegbuf)
         xint_counter = integrate_energy(p.wn, p.wt, p.we, wd, -wb, nueff,
                                              p.ell, leff, p.n, p.psi, lambda, p.method;
                                              nutype=p.nutype, f0type=p.f0type,
                                              nufac=p.nufac, ximag=p.ximag, qt=p.qt,
-                                             atol=p.energy_atol, rtol=p.energy_rtol)
+                                             atol=p.energy_atol, rtol=p.energy_rtol, segbuf=p.esegbuf)
         xint = xint_co + xint_counter
     else
         xint = integrate_energy(p.wn, p.wt, p.we, wd, wb, nueff,