# Reproduction of NuFIT 6.0, JHEP 12 (2024) 216 [arXiv:2410.05380], Fig. 11: solar vs KamLAND determination of
# θ₁₂ and Δm²₂₁, with Newtrinos.jl (solar modules `chlorine`, `gallex_gno`, `sage`, `sno`, `sk1_solar`–`sk4_solar`,
# `borexino_ph1`–`borexino_ph3`; `kamland`). Borexino enters through its published interaction rates.
# Run from this directory with the Newtrinos.jl environment at the pinned commit:
#   julia -t 64 --project=<path to Newtrinos.jl> reproduce.jl
using Newtrinos, Distributions, DataStructures, FileIO, CairoMakie, CSV, DataFrames, LinearAlgebra, Interpolations

cd(@__DIR__); mkpath("ours"); mkpath("results"); mkpath("cache")
BLAS.set_num_threads(1)

# CSV export of a profile scan: one row per grid point, ASCII column names, oscillation parameters in physical
# units (sin²θ, Δm² in eV²), dchi2 = -2Δlog L w.r.t. `ref` (default: the scan's best fit)
function save_csv(path, r; ref = maximum(r.values.log_posterior))
    ascii(k) = replace(string(k), "θ" => "theta", "Δm²" => "dm2_", "δ" => "d", "₁" => "1", "₂" => "2", "₃" => "3", "CP" => "cp")
    cols = OrderedDict{String,Any}()
    for (k, v) in pairs(r.values)
        x = vec(v)
        startswith(string(k), "θ") && (cols["sin2_" * ascii(k)] = sin.(x) .^ 2)
        cols[ascii(k)] = x
    end
    cols["dchi2"] = 2 .* (ref .- vec(r.values.log_posterior))
    haskey(cols, "llh") && (cols["log_likelihood"] = pop!(cols, "llh"))
    first_cols = filter(in(keys(cols)), ["sin2_theta12", "dm2_21", "dchi2", "log_likelihood", "log_posterior"])
    CSV.write(path, DataFrame(cols)[:, vcat(first_cols, setdiff(collect(keys(cols)), first_cols))])
end

# NuFIT Fig. 11 conditions: sin²θ₁₃ = 0.0222 fixed; θ₂₃, δCP, Δm²₃₁ irrelevant and fixed; all flux, cross-section and
# detector nuisance parameters profiled with their priors
function setup(experiments)
    p = Newtrinos.get_params(experiments)
    priors = Newtrinos.get_priors(experiments)
    θ13 = asin(sqrt(0.0222))
    p = merge(p, (θ₁₃ = θ13, Δm²₂₁ = 6e-5))
    priors = merge(priors, (θ₁₃ = θ13, θ₂₃ = p.θ₂₃, δCP = p.δCP, Δm²₃₁ = p.Δm²₃₁,
                            θ₁₂ = Uniform(asin(sqrt(0.18)), asin(sqrt(0.42))), Δm²₂₁ = Uniform(1.5e-5, 1.4e-4)))
    Newtrinos.generate_likelihood(experiments), priors, p
end

physics = Newtrinos.solar_common.default_physics()
# SK day/night information from SK's amplitude fit of the zenith-angle variation (SK-I–IV, sk_solar_dn), with the SK
# day/night spectra merged per energy bin
solar = merge(NamedTuple(name => (name in (:sk1_solar, :sk2_solar, :sk3_solar, :sk4_solar) ?
                                  getproperty(Newtrinos, name).configure(physics; daynight = :combined) :
                                  getproperty(Newtrinos, name).configure(physics))
                         for name in (:chlorine, :gallex_gno, :sage, :sno, :sk1_solar, :sk2_solar, :sk3_solar, :sk4_solar,
                                      :borexino_ph1, :borexino_ph2, :borexino_ph3)),
              (sk_solar_dn = Newtrinos.sk_solar_dn.configure(physics; dataset = :sk1to4),))
kamland = (kamland = Newtrinos.kamland.configure(),)

llh, priors, p = setup(solar)
# coarse grids: this entry validates the setup, it does not aim at smooth publication-quality contours
solar_2d = Newtrinos.profile(llh, priors, OrderedDict(:θ₁₂ => 10, :Δm²₂₁ => 10), p, cache_dir = "cache/solar_2d")
priors_dm = merge(priors, (Δm²₂₁ = Uniform(2e-5, 1e-4),))
solar_dm = Newtrinos.profile(llh, priors_dm, OrderedDict(:Δm²₂₁ => 17), p, cache_dir = "cache/solar_dm")
llh, priors, p = setup(kamland)
kamland_dm = Newtrinos.profile(llh, merge(priors, (Δm²₂₁ = Uniform(2e-5, 1e-4),)), OrderedDict(:Δm²₂₁ => 161), merge(p, (Δm²₂₁ = 7.5e-5,)),
                               cache_dir = "cache/kamland_dm")

FileIO.save("results/nufit6_fig11.jld2", Dict("solar_th12_dm21" => solar_2d, "solar_dm21" => solar_dm, "kamland_dm21" => kamland_dm))
save_csv("results/solar_th12_dm21.csv", solar_2d)
save_csv("results/solar_dm21.csv", solar_dm)
save_csv("results/kamland_dm21.csv", kamland_dm)

Δχ²(r) = 2 .* (maximum(r.values.log_posterior) .- r.values.log_posterior)

fig = Figure(size = (900, 520))
# left: solar regions, 1σ–5σ (2 dof), and KamLAND best fit
ax = Axis(fig[1, 1], title = "Newtrinos solar (Cl, Ga, SK I–IV, SNO, Borexino), sin²θ₁₃ = 0.0222",
          xlabel = "sin²θ₁₂", ylabel = "Δm²₂₁ (10⁻⁵ eV²)", titlesize = 13)
x = sin.(solar_2d.axes.θ₁₂) .^ 2; y = solar_2d.axes.Δm²₂₁ .* 1e5
levels = [quantile(Chisq(2), 1 - 2 * ccdf(Normal(), n)) for n in 1:5]
contourf!(ax, x, y, Δχ²(solar_2d), levels = vcat(0, levels), colormap = [:red, :pink, :blue, :magenta, :cyan])
contour!(ax, x, y, Δχ²(solar_2d), levels = levels, color = :black, linewidth = 0.5)
i = argmax(solar_2d.values.log_posterior)
scatter!(ax, [x[i[1]]], [y[i[2]]], color = :white, strokecolor = :black, strokewidth = 1)
kl_best = kamland_dm.axes.Δm²₂₁[argmax(kamland_dm.values.log_posterior)]
hlines!(ax, [kl_best * 1e5], color = :green, linestyle = :dash, label = "KamLAND best fit Δm²₂₁")
xlims!(ax, 0.2, 0.4); ylims!(ax, 0, 14)
axislegend(ax, position = :rt, labelsize = 11)
# right: Δχ²(Δm²₂₁), θ₁₂ profiled
ax2 = Axis(fig[1, 2], xlabel = "Δm²₂₁ (10⁻⁵ eV²)", ylabel = "Δχ²", title = "θ₁₂ and nuisance parameters profiled", titlesize = 13)
lines!(ax2, solar_dm.axes.Δm²₂₁ .* 1e5, Δχ²(solar_dm), color = :red, label = "Solar (B23 MB22-met)")
lines!(ax2, kamland_dm.axes.Δm²₂₁ .* 1e5, Δχ²(kamland_dm), color = :green, label = "KamLAND")
hlines!(ax2, [1, 4, 9], color = :gray, linewidth = 0.5)
xlims!(ax2, 2, 10); ylims!(ax2, 0, 12)
axislegend(ax2, position = :rt, labelsize = 11)
save("ours/fig11_solar_kamland.png", fig)

# key numbers for the summary
# cubic spline through the coarse 1D solar profile
dm_ax = range(first(solar_dm.axes.Δm²₂₁), last(solar_dm.axes.Δm²₂₁), length = length(solar_dm.axes.Δm²₂₁))
@assert collect(dm_ax) ≈ solar_dm.axes.Δm²₂₁
χ²_solar = cubic_spline_interpolation(dm_ax, Δχ²(solar_dm))
dm_fine = range(first(dm_ax), last(dm_ax), length = 2001)
best = dm_fine[argmin(χ²_solar.(dm_fine))]
at_kl = χ²_solar(kl_best) - minimum(χ²_solar.(dm_fine))
open("results/summary.txt", "w") do io
    println(io, "solar best-fit sin2_theta12 = ", round(x[i[1]], digits = 3), ", dm2_21 = ", round(y[i[2]], digits = 2), "e-5 eV^2 (2D grid)")
    println(io, "solar best-fit dm2_21 (1D profile) = ", round(best * 1e5, digits = 2), "e-5 eV^2")
    println(io, "KamLAND best-fit dm2_21 = ", round(kl_best * 1e5, digits = 2), "e-5 eV^2")
    println(io, "solar dchi2 at the KamLAND best fit = ", round(at_kl, digits = 2))
end
print(read("results/summary.txt", String))

# ═══════════════════════════════════════════════════════════════════════════════════════════════════════════════════
# Fig. 1: global 3ν fit, variant «IC19 w/o SK-atm». Newtrinos approximation of the NuFIT data set: KamLAND, Daya Bay,
# T2K, NOvA, MINOS and IceCube DeepCore 3 y in full, plus all solar data as a profiled Δχ² table in
# (θ₁₂, Δm²₂₁, sin²θ₁₃). 1D profiles of the six oscillation parameters for each ordering, every point minimised over all
# other parameters from two starting points (θ₂₃ octants; δCP for the θ₂₃ profile). The fits run on NPROC worker
# processes × NTHR threads (env, default 24 × 5).
using Distributed, Printf
const NPROC = parse(Int, get(ENV, "NPROC", "24")); const NTHR = parse(Int, get(ENV, "NTHR", "5"))
"Run `f` over `jobs` in batches of NTHR (threads within a worker), on NPROC freshly started workers set up by `init`."
function distributed_fits(init, jobs)
    ws = addprocs(NPROC; exeflags = `-t $NTHR --project=$(Base.active_project())`)
    try
        Distributed.remotecall_eval(Main, ws, init)
        batches = [jobs[i:min(i + NTHR - 1, end)] for i in 1:NTHR:length(jobs)]
        reduce(vcat, pmap(b -> Main.fit_batch(b), WorkerPool(ws), batches))
    finally
        rmprocs(ws)
    end
end
const ENTRY = @__DIR__

# solar Δχ² table: Cl, GALLEX/GNO, SAGE, SNO, SK-I–IV (day/night spectra), Borexino Phase I (rates), Phase II
# (spectral fit), Phase III; solar nuisance parameters profiled; θ₂₃, δCP, Δm²₃₁ irrelevant and fixed
solar_init = quote
    using Newtrinos, Distributions, LinearAlgebra
    using BAT: distprod
    BLAS.set_num_threads(1)
    const sp = Newtrinos.solar_common.default_physics()
    const solar = (chlorine = Newtrinos.chlorine.configure(sp), gallex_gno = Newtrinos.gallex_gno.configure(sp),
                   sage = Newtrinos.sage.configure(sp), sno = Newtrinos.sno.configure(sp), sk_solar = Newtrinos.sk_solar.configure(sp),
                   borexino_ph1 = Newtrinos.borexino_ph1.configure(sp),
                   borexino_ph2_spectrum = Newtrinos.borexino_ph2_spectrum.configure(sp),
                   borexino_ph3 = Newtrinos.borexino_ph3.configure(sp))
    const LLH = Newtrinos.generate_likelihood(solar)
    const P0 = merge(Newtrinos.get_params(solar), (θ₂₃ = asin(sqrt(0.47)), δCP = 3.70, Δm²₃₁ = 2.513e-3))
    const PR0 = merge(Newtrinos.get_priors(solar), (θ₂₃ = P0.θ₂₃, δCP = P0.δCP, Δm²₃₁ = P0.Δm²₃₁))
    function fit_point(θ12, dm21, θ13)
        p = merge(P0, (θ₁₂ = θ12, Δm²₂₁ = dm21, θ₁₃ = θ13))
        pr = merge(PR0, (θ₁₂ = θ12, Δm²₂₁ = dm21, θ₁₃ = θ13))
        Newtrinos.find_mle_cached(LLH, distprod(; pr...), p, joinpath($ENTRY, "cache", "solar_table");
                                  adsel = Newtrinos.ADTypes.AutoForwardDiff())[2]
    end
    fit_batch(pts) = (out = Vector{Float64}(undef, length(pts)); Threads.@threads for i in eachindex(pts); out[i] = fit_point(pts[i]...); end; out)
end
θ12_tab = collect(range(asin(sqrt(0.20)), asin(sqrt(0.42)), length = 15))
dm21_tab = collect(range(4.0e-5, 11.0e-5, length = 21))
pts = vec([(a, b, asin(sqrt(c))) for a in θ12_tab, b in dm21_tab, c in [0.0195, 0.0222, 0.0249]])
mkpath("cache/solar_table")
lp_solar = distributed_fits(solar_init, pts)
CSV.write("results/solar_table.csv", DataFrame(sin2_theta12 = [sin(p[1])^2 for p in pts], dm2_21 = [p[2] for p in pts],
                                               sin2_theta13 = [sin(p[3])^2 for p in pts], minus2lnL = -2 .* lp_solar))

global_init = quote
    using Newtrinos, Distributions, LinearAlgebra, CSV, DataFrames
    using BAT: distprod
    BLAS.set_num_threads(1)
    "Solar Δχ² table as a chi2map experiment, interpolated in (θ₁₂, Δm²₂₁, sin²θ₁₃)."
    function solar_term(file)
        df = CSV.read(file, DataFrame)
        θ = sort(unique(asin.(sqrt.(df.sin2_theta12)))); dm = sort(unique(df.dm2_21)); s13 = sort(unique(df.sin2_theta13))
        tab = fill(NaN, length(θ), length(dm), length(s13))
        for r in eachrow(df)
            tab[searchsortedfirst(θ, asin(sqrt(r.sin2_theta12)) - 1e-12), searchsortedfirst(dm, r.dm2_21 - 1e-20),
                searchsortedfirst(s13, r.sin2_theta13 - 1e-12)] = r.minus2lnL
        end
        @assert !any(isnan, tab)
        ax = (range(θ[1], θ[end], length = length(θ)), range(dm[1], dm[end], length = length(dm)), range(s13[1], s13[end], length = length(s13)))
        Newtrinos.chi2map.configure((p -> p.θ₁₂, p -> p.Δm²₂₁, p -> sin(p.θ₁₃)^2), ax, tab; title = "solar (all data, profiled)")
    end
    "Experiments, likelihood, start parameters and priors (flat in θ, δCP, Δm²) for one mass ordering."
    function setup(ordering)
        exps = (kamland = Newtrinos.kamland.configure(), dayabay = Newtrinos.dayabay.configure(),
                t2k = Newtrinos.t2k.configure(), nova = Newtrinos.nova.configure(), minos = Newtrinos.minos.configure(),
                deepcore_3y = Newtrinos.deepcore_3y.configure(), solar = solar_term(joinpath($ENTRY, "results", "solar_table.csv")))
        p = Newtrinos.get_params(exps)
        pr = Newtrinos.get_priors(exps)
        if ordering == :NO
            p = merge(p, (Δm²₃₁ = 2.51e-3,)); pr = merge(pr, (Δm²₃₁ = Uniform(2.2e-3, 2.9e-3),))
        else
            p = merge(p, (Δm²₃₁ = -2.41e-3,)); pr = merge(pr, (Δm²₃₁ = Uniform(-2.8e-3, -2.1e-3),))
        end
        p = merge(p, (θ₁₂ = asin(sqrt(0.307)), θ₁₃ = asin(sqrt(0.0222)), θ₂₃ = asin(sqrt(0.5)), δCP = 3.7, Δm²₂₁ = 7.5e-5))
        pr = merge(pr, (θ₁₂ = Uniform(asin(sqrt(0.24)), asin(sqrt(0.38))), θ₁₃ = Uniform(asin(sqrt(0.018)), asin(sqrt(0.026))),
                        θ₂₃ = Uniform(asin(sqrt(0.35)), asin(sqrt(0.65))), δCP = Uniform(0, 2π), Δm²₂₁ = Uniform(6.0e-5, 9.0e-5)))
        (; exps, llh = Newtrinos.generate_likelihood(exps), p, pr)
    end
    const SETUPS = Dict(o => setup(o) for o in (:NO, :IO))
    "Profile log likelihood with `var` fixed to `val`: the log posterior minus the (constant) flat oscillation priors."
    function fit_point(ordering, var, val, start)
        s = SETUPS[ordering]
        p = merge(s.p, start, NamedTuple{(var,)}((val,)))
        pr = merge(s.pr, NamedTuple{(var,)}((val,)))
        r = Newtrinos.find_mle_cached(s.llh, distprod(; pr...), p, joinpath($ENTRY, "cache", "global");
                                      adsel = Newtrinos.ADTypes.AutoForwardDiff())
        r[2] - sum(logpdf(pr[k], r[3][k]) for k in (:θ₁₂, :θ₁₃, :θ₂₃, :δCP, :Δm²₂₁, :Δm²₃₁) if k != var)
    end
    fit_batch(jobs) = (out = Vector{Float64}(undef, length(jobs)); Threads.@threads for i in eachindex(jobs); out[i] = fit_point(jobs[i]...); end; out)
end
s2θ(x) = asin(sqrt(x))
function grid(ordering)
    dm3l = ordering == :NO ? range(2.40e-3, 2.64e-3, length = 17) : range(-2.62e-3, -2.38e-3, length = 17) .+ 7.5e-5   # Δm²₃₁ (IO: Δm²₃₂ + Δm²₂₁)
    (θ₁₂ = s2θ.(range(0.26, 0.36, length = 15)), θ₁₃ = s2θ.(range(0.0195, 0.0250, length = 15)),
     θ₂₃ = s2θ.(range(0.40, 0.62, length = 23)), δCP = deg2rad.(0:15:345), Δm²₂₁ = range(6.6e-5, 8.4e-5, length = 15),
     Δm²₃₁ = dm3l)
end
starts(var) = var == :θ₂₃ ? [(δCP = 3.4,), (δCP = 4.7,)] : [(θ₂₃ = s2θ(0.45),), (θ₂₃ = s2θ(0.57),)]
jobs = [(o, var, v, st) for o in (:NO, :IO) for var in keys(grid(o)) for v in grid(o)[var] for st in starts(var)]
mkpath("cache/global")
lp_global = distributed_fits(global_init, jobs)
raw = DataFrame(ordering = [string(j[1]) for j in jobs], var = [string(j[2]) for j in jobs], value = [j[3] for j in jobs],
                log_likelihood = lp_global)
prof = combine(groupby(raw, [:ordering, :var, :value]), :log_likelihood => maximum => :log_likelihood)
prof.dchi2 = 2 .* (maximum(prof.log_likelihood) .- prof.log_likelihood)    # w.r.t. the best fit of both orderings
# plotted quantities as in NuFIT: sin²θ, δCP in degrees, Δm²₂₁ in 10⁻⁵ eV², Δm²₃₁ (NO) and Δm²₃₂ (IO) in 10⁻³ eV²
# (Δm²₃₂ = Δm²₃₁ − Δm²₂₁ with Δm²₂₁ = 7.5×10⁻⁵ eV²; the shift from the fitted Δm²₂₁ is < 10⁻⁶ eV²)
xplot(var, o, v) = var == "Δm²₃₁" ? (o == "NO" ? v : v - 7.5e-5) * 1e3 : var == "δCP" ? rad2deg(v) :
                   var == "Δm²₂₁" ? v * 1e5 : sin(v)^2
prof.x = [xplot(r.var, r.ordering, r.value) for r in eachrow(prof)]
sort!(prof, [:ordering, :var, :x])
CSV.write("results/global_profiles.csv", prof[:, [:ordering, :var, :x, :dchi2, :log_likelihood]])

tag = Dict("θ₁₂" => "T12", "θ₁₃" => "T13", "θ₂₃" => "T23", "δCP" => "DCP", "Δm²₂₁" => "DMS", "Δm²₃₁" => "DMA")
function nufit(o, var)
    t = CSV.read("data/nufit6_TBoff/TBoff_$(o)_$(tag[var]).csv", DataFrame)
    x = var == "Δm²₂₁" ? 10 .^ t.c0 .* 1e5 : var == "δCP" ? mod.(t.c0, 360) : t.c0
    k = sortperm(x); x, y = x[k], t.c1[k]
    k = unique(i -> x[i], eachindex(x)); x[k], y[k]               # δCP: 0° and 360° both tabulated
end
ours(o, var) = (d = prof[(prof.ordering .== o) .& (prof.var .== var), :]; (d.x, d.dchi2))

# layout of NuFIT Fig. 1
fig = Figure(size = (900, 1200))
panels = [("θ₁₂", 1, 1, "sin²θ₁₂", (0.2, 0.42)), ("Δm²₂₁", 1, 2, "Δm²₂₁ [10⁻⁵ eV²]", (6.3, 8.5)),
          ("θ₂₃", 2, 1, "sin²θ₂₃", (0.38, 0.65)), ("θ₁₃", 3, 1, "sin²θ₁₃", (0.018, 0.0265)), ("δCP", 3, 2, "δCP [°]", (0, 360))]
style(o) = o == "NO" ? :red : :blue
for (var, r, c, lab, xr) in panels
    ax = Axis(fig[r, c], xlabel = lab, ylabel = "Δχ²")
    for o in ("NO", "IO")
        lines!(ax, nufit(o, var)..., color = (style(o), 0.4), linewidth = 5)
        lines!(ax, ours(o, var)..., color = style(o), linewidth = 1.5)
    end
    var == "δCP" && (ax.xticks = 0:90:360)
    hlines!(ax, [1, 4, 9], color = :gray80, linewidth = 0.5); xlims!(ax, xr...); ylims!(ax, 0, 16)
end
g = fig[2, 2] = GridLayout()
for (k, o, xr, lab) in ((1, "IO", (-2.66, -2.34), "Δm²₃₂ [10⁻³ eV²]"), (2, "NO", (2.34, 2.66), "Δm²₃₁ [10⁻³ eV²]"))
    ax = Axis(g[1, k], xlabel = lab, ylabel = k == 1 ? "Δχ²" : "", yticklabelsvisible = k == 1)
    lines!(ax, nufit(o, "Δm²₃₁")..., color = (style(o), 0.4), linewidth = 5)
    lines!(ax, ours(o, "Δm²₃₁")..., color = style(o), linewidth = 1.5)
    hlines!(ax, [1, 4, 9], color = :gray80, linewidth = 0.5); xlims!(ax, xr...); ylims!(ax, 0, 16)
end
colgap!(g, 4)
Legend(fig[0, :], [LineElement(color = (:red, 0.4), linewidth = 5), LineElement(color = (:blue, 0.4), linewidth = 5),
                   LineElement(color = :red, linewidth = 1.5), LineElement(color = :blue, linewidth = 1.5)],
       ["NuFIT 6.0 NO", "NuFIT 6.0 IO", "Newtrinos NO", "Newtrinos IO"], orientation = :horizontal, framevisible = false)
Label(fig[-1, :], "Global 3ν fit, «IC19 w/o SK-atm» (Newtrinos: KamLAND, Daya Bay, T2K, NOvA, MINOS, DeepCore 3 y, solar)", fontsize = 14)
save("ours/fig1_global_profiles.png", fig)

# best fits and 1σ ranges (Δχ² = 1 w.r.t. each ordering's minimum), as in NuFIT Table 1
function onesigma(x, y)
    m = minimum(y); itp = linear_interpolation(x, y .- m)
    fine = range(first(x), last(x), length = 4001); inside = fine[itp.(fine) .<= 1]
    x[argmin(y)], first(inside), last(inside), m
end
open("results/summary_global.txt", "w") do io
    println(io, "best fit [1σ range] (Δχ² of the ordering's minimum); Newtrinos | NuFIT 6.0 IC19 w/o SK-atm")
    for var in ("θ₁₂", "θ₁₃", "θ₂₃", "δCP", "Δm²₂₁", "Δm²₃₁"), o in ("NO", "IO")
        f(t) = @sprintf("%.4g [%.4g, %.4g] (%.2f)", t...)
        println(io, rpad("$(tag[var]) $o", 8), rpad(f(onesigma(ours(o, var)...)), 40), "| ", f(onesigma(nufit(o, var)...)))
    end
end
print(read("results/summary_global.txt", String))

# ═══════════════════════════════════════════════════════════════════════════════════════════════════════════════════
# Fig. 8: (δCP, sin²θ₂₃) regions for IO and NO w.r.t. the global best fit, all other parameters profiled (same global
# fit as Fig. 1); coarse 12 × 12 grid per ordering
dcp_2d = deg2rad.(0:30:330); s23_2d = range(0.40, 0.62, length = 12)
jobs2d = [(o, (δCP = d, θ₂₃ = s2θ(s)), nothing) for o in (:NO, :IO) for s in s23_2d for d in dcp_2d]
fit2d_init = quote
    $global_init
    "Profile log likelihood with θ₂₃ and δCP fixed."
    function fit_point(ordering, fixed, _)
        s = SETUPS[ordering]
        r = Newtrinos.find_mle_cached(s.llh, distprod(; merge(s.pr, fixed)...), merge(s.p, fixed), joinpath($ENTRY, "cache", "global_2d");
                                      adsel = Newtrinos.ADTypes.AutoForwardDiff())
        r[2] - sum(logpdf(s.pr[k], r[3][k]) for k in (:θ₁₂, :θ₁₃, :Δm²₂₁, :Δm²₃₁))
    end
end
mkpath("cache/global_2d")
lp2d = distributed_fits(fit2d_init, jobs2d)
gbest = max(maximum(prof.log_likelihood), maximum(lp2d))
CSV.write("results/global_dcp_th23.csv", DataFrame(ordering = [string(j[1]) for j in jobs2d], sin2_theta23 = [sin(j[2].θ₂₃)^2 for j in jobs2d],
                                                   dcp_deg = [rad2deg(j[2].δCP) for j in jobs2d], dchi2 = 2 .* (gbest .- lp2d)))

levels2 = [2.30, 4.61, 6.18, 9.21, 11.83]        # 1σ, 90 %, 2σ, 99 %, 3σ (2 dof)
fig = Figure(size = (1000, 480))
for (k, o) in enumerate(("IO", "NO"))
    ax = Axis(fig[1, k], xlabel = "δCP [°]", ylabel = k == 1 ? "sin²θ₂₃" : "", title = o, xticks = 0:90:360)
    # NuFIT: filled regions from the 2D table (δCP −180…180 → 0…360)
    t = CSV.read("data/nufit6_TBoff/TBoff_$(o)_T23_DCP.csv", DataFrame)
    sx = sort(unique(t.c0)); dx = sort(unique(mod.(t.c1, 360)))
    z = fill(NaN, length(dx), length(sx))
    for r in eachrow(t); z[searchsortedfirst(dx, mod(r.c1, 360)), searchsortedfirst(sx, r.c0)] = r.c2; end
    contourf!(ax, dx, sx, z, levels = vcat(0, levels2), colormap = [:red, :pink, :blue, :magenta, :cyan])
    # Newtrinos: contours of the coarse grid, bicubic in (δCP periodic, sin²θ₂₃)
    zn = [2 * (gbest - lp2d[findfirst(j -> j[1] == Symbol(o) && j[2].δCP == d && j[2].θ₂₃ == s2θ(s), jobs2d)]) for d in dcp_2d, s in s23_2d]
    itp = cubic_spline_interpolation((range(0, 2π, length = 13), s23_2d), vcat(zn, zn[1:1, :]))
    df = range(0, 2π, length = 181); sf = range(first(s23_2d), last(s23_2d), length = 111)
    contour!(ax, rad2deg.(df), sf, [itp(a, b) for a in df, b in sf], levels = levels2, color = :black, linewidth = 1.2)
    i = argmin(zn); scatter!(ax, [rad2deg(dcp_2d[i[1]])], [s23_2d[i[2]]], color = :black, marker = :xcross)
    xlims!(ax, 0, 360); ylims!(ax, 0.35, 0.65)
end
Label(fig[0, :], "«IC19 w/o SK-atm», regions w.r.t. the global best fit (1σ, 90%, 2σ, 99%, 3σ, 2 dof): NuFIT 6.0 filled, Newtrinos black lines", fontsize = 13)
save("ours/fig8_dcp_th23.png", fig)