v0.2.0

Added

• AMR checkpoint/restart (ADC-65): AmrSystem.checkpoint(path) / restart(path) write and read
  a bit-identical npz (per-level conservative state with fine patches, per-level phi as the multigrid
  warm-start, the fine hierarchy and the clock), replacing the old NotImplementedError. Restart
  imposes the saved fine layout via the new set_hierarchy (no re-clustering, no re-prolongation).
  Mono-block mono-rank for now: MPI np>1, multi-block and regrid_every>0 are rejected explicitly.
  Adds the append-only level accessors n_levels / n_vars / level_state / level_potential /
  set_hierarchy.

• Model-declared named aux fields (ADC-70, phase 1, Cartesian System): a model declares
  persistent auxiliary fields with m.aux_field("name") (read in a formula via aux.extra_field(k));
  sim.set_aux_field(block, name, array) and sim.aux_field(block, name) set and read them (up to
  four). The Aux POD stays trivially copyable and device-clean, named components are static (never
  rewritten by solve_fields), and a model with no named field is cache/hash-identical to before;
  B_z / T_e stay on their dedicated setters.

• Generic 2D moment-hierarchy generator (ADC-164): the new adc.moments module derives the full
  M->C->S->closure->C'->M' algebra (binomial transform plus standardization) over the DSL AST, so a
  user supplies only the closure (S -> standardized moments of order N+1). Ships build_moment_model,
  gaussian_closure (Levermore), lorentz_sources (Vlasov-Lorentz) and moment_indices /
  moment_names; robust=True adds differentiable smooth floors and exact_speeds=True wires HLL
  speeds via autodiff plus numeric eig.

• SSPRK3 time integration on AMR (ADC-64): AmrSystem.add_block(..., time="ssprk3") (or
  adc.Explicit(ssprk3=True)) runs a 3-stage Shu-Osher SSPRK3 per subcycled level, staying exactly
  conservative across coarse-fine boundaries by refluxing the effective flux
  Feff = 1/6 F0 + 1/6 F1 + 2/3 F2. The default stays forward Euler (bit-identical); rejected
  explicitly on the compiled .so paths and with imex.

• First-order forward Euler explicit method (ADC-174): adc.Explicit(method="euler") selects the
  ForwardEuler stepper on System.add_block (native and backend="production"), for
  first-order-reference fidelity; ssprk2 stays the default. The frozen-ABI AOT path rejects euler
  (pointing to the production/native backends) rather than silently ignoring it.

• Explicit signed wave speeds in the DSL (ADC-83): m.wave_speeds(x=(smin,smax), y=(smin,smax))
  declares signed face speeds directly, unlocking riemann="hll" for models with no pressure
  primitive (moment systems, isothermal). Without set_eigenvalues the Rusanov/CFL bound derives
  from max(|smin|,|smax|); an add_equation guard rejects riemann="hll" with no emitted speeds.

• HLL wave speeds from the flux Jacobian (ADC-86/87): m.wave_speeds_from_jacobian(x=, y=, eig="numeric"|"fd", blocks=) emits exact HLL speeds as the spectrum extremes of the flux Jacobian
  (new device-clean header adc::real_eig_minmax: closed form for N<=2, Hessenberg plus Francis QR
  otherwise, Gershgorin outer-bound fallback). x/y=None autodiffs the declared flux; the Abs
  autodiff node is now differentiable, so smooth max(x,eps) floors are Jacobian-able.

• Positivity floor limiter (ADC-76): adc.FiniteVolume(positivity_floor=...) imposes a density
  floor on reconstructed face states, falling back locally to the source-cell average (conservative,
  first-order) on a violating face. floor<=0 is bit-identical to before; it is threaded through the
  Cartesian, polar and cut-cell kernels and the compiled-block ABI, and rejected explicitly on the
  paths that do not support it (prototype JIT, AOT/production, AMR).

• Opt-in HLL wave-speed cache (ADC-199): adc.FiniteVolume(wave_speed_cache=True) evaluates
  model.wave_speeds once per cell and direction and reuses it as the face bound instead of
  recomputing on every face and RK stage (measured speedup on moment hierarchies). Off by default,
  bit-identical to OFF in the no-slope path, and rejected outside riemann="hll" plus explicit time.

• Spectral Poisson FFT variant (ADC-175): a new elliptic kind solver="fft_spectral" (from
  System.set_poisson, adc.EllipticSolver, listed in adc.capabilities()) reuses the periodic
  single-rank FFT plumbing of "fft" but with the continuous Laplacian symbol -(kx^2+ky^2), exact
  on sinusoids. The discrete "fft" and "geometric_mg" solvers are unchanged.

• Native Windows (MSVC) support (ADC-99/100/136/144): adc_cpp compiles and imports natively on
  Windows without WSL2. A portable dynamic-loading layer adc::dynlib (LoadLibraryW /
  GetProcAddress), an ADC_EXPORT macro (__declspec), an MSVC-aware ABI key, and
  std::numbers::pi for M_PI cover the runtime; the DSL production backend compiles a model to a
  .dll with cl and runs bit-identical to the brick path (shared Kokkos, _adc.lib). All _WIN32
  branches are dead off Windows.

• System.dt_hotspot(name) CFL diagnostic (ADC-182): returns {w, i, j} for the global cell
  that dominates a block's transport CFL bound, to locate a collapsing dt without an external scan.
  On-demand and off the hot path (step / step_cfl stay bit-identical), a two-pass device-clean
  reduction with an MPI all-reduce.

• Find-or-fetch Kokkos (ADC-263): if no Kokkos install is found, CMake downloads and builds a
  release tarball verified by SHA256, so a plain cmake -B build works without a pre-installed
  Kokkos. Overridable via ADC_KOKKOS_FETCH_VERSION (default 4.4.01) and ADC_KOKKOS_FETCH_SHA256.

• Eigenvalue witness in the projection DSL (ADC-289): dsl.eig_max_im(rows), dsl.eig_lmin(rows)
  and dsl.eig_lmax(rows) build a small dense matrix from moment expressions and return a scalar
  Expr from its spectrum via adc::real_eig_minmax (max imaginary part as a complex-eigenvalue
  witness, or the real-part bounds). The codegen emits a named device-clean functor (no extended
  lambda), so m.projection can express a branchless "if a moment matrix has a complex eigenvalue,
  correct it" rule (unblocks the native relaxation15 projector, ADC-275). Additive: the existing
  expression set and m.projection (ADC-177) are unchanged.

• Documentation rebuilt around a Diataxis navigation (ADC-248): Getting started, Tutorials,
  Concepts (11 pages, ADC-249), How-to (12 task pages), Simulation, AMR, Running, Advanced topics,
  Reference and Development sections, plus a Quickstart page and an internal documentation style
  guide (ADC-250). Pages are kebab-case.

• Embedded C++ reference in the Sphinx site via doxysphinx (ADC-149), with a modern Doxygen
  theme and full dot diagrams (ADC-239).

• Documentation-as-code tooling: per-page docmap.toml with owner and freshness plus an
  example harness (ADC-147), a doc taxonomy and a stack ADR (ADC-146), and CI doc lanes (a light PR
  lane and a weekly heavy lane with lld and ccache, ADC-151/225).

• Quality tooling / static analysis (ADC-105): dedicated CI workflow .github/workflows/quality.yml,
  off the PR critical path (weekly Sunday cron + workflow_dispatch + quality label). Five
  informative (non-blocking) jobs: clang-format (.clang-format), strict warnings
  (ADC_ENABLE_WARNINGS), clang-tidy (.clang-tidy), ASan+UBSan sanitizers (ADC_ENABLE_SANITIZERS,
  ci-warnings/ci-asan presets) and CodeQL. CMake options OFF by default (empty adc_dev_options
  target), so ci.yml, local builds and adc_cases are unchanged. See docs/QUALITY_TOOLING.md.

• Fuzzing, coverage and developer automation (ADC-113): invariant-checked libFuzzer harnesses in
  fuzz/ (Box2D, Berger-Rigoutsos clustering, real_eig_minmax; option ADC_BUILD_FUZZING +
  ci-fuzz preset, clang), gcov/gcovr coverage (ADC_ENABLE_COVERAGE + ci-coverage preset),
  Python ruff lint ([tool.ruff]), opt-in pre-commit hooks (.pre-commit-config.yaml: clang-format
  and ruff at commit), and two more quality.yml jobs (fuzz, coverage) on the same weekly
  informative cadence; the default build stays bit-unchanged.

• Repository health and GitHub hygiene: BSD-3 LICENSE and license declaration, CONTRIBUTING,
  SECURITY, .gitattributes, PR and issue templates (ADC-223/224/244/246), a root-hygiene guard
  (ADC-169), Dependabot weekly Actions bumps (ADC-117), a release workflow that turns a v* tag
  into a GitHub Release from this changelog (ADC-234), and docs/VERSIONING.md with the bump rules
  (ADC-232/237).

• Shared C++ test harness (test_harness.hpp, bench/common.hpp, ADC-215); a Schur-split test
  with an AMR guard on add_block and auto-skip without Kokkos (ADC-207).

• CI: ci.yml split into parallel gate-cpp / gate-python lanes with mold and path
  routing (ADC-171).

• Moment-model documentation: a step-by-step HyQMOM tutorial, a moments-and-closures concept
  page, a moment-models reference, and adc.moments (build_moment_model, gaussian_closure,
  lorentz_sources) added to the Python API reference (autodoc).

• Generic pointwise post-step PROJECTION hook (ADC-177): m.projection([...]) on the DSL side
  (C++ trait HasPointwiseProjection, compiled like the flux/source), applied by System at the
  end of every whole macro-step (never per RK stage) on the valid cells of each block -- replaces
  the per-cell Python callback (aot and production backends; prototype and
  target="amr_system" reject it explicitly). Adds dsl.sign(x) (branch-free mask selections,
  differentiable). See docs/DSL_API.md section 5.

Changed

• Kokkos is the only on-node backend (ADC-263): the standalone OpenMP path is removed (the
  ADC_USE_OPENMP option, find_package(OpenMP), the mutual-exclusion check and every #pragma omp
  or manual h...