feat(fdtd): add multi-GPU joint solve

FEATURE_LIST.md

@@ -233,6 +233,9 @@ result = mw.Simulation.fdtd(scene, frequencies=[200e12]).run()
 - Time-domain simulation through `Simulation.fdtd(...)`
 - CUDA-only solver execution; `Simulation.fdtd(...)` requires `Scene(device="cuda")`
 - Compiled native CUDA kernels are the only FDTD runtime backend; the Torch reference path has been retired from the runtime and relocated to the CUDA parity tests
+- Engineering-preview multi-GPU joint solves through public `FDTDParallelConfig` and `Simulation.fdtd(..., parallel=...)`: Python remains the control plane, and `DistributedFDTD` owns an independent distributed coordinator/time loop for x-slab ownership, P2P halos, streams, and events. Single- and multi-GPU execution share the same native CUDA numerical core; the six bounded standard real-field operations and legacy full-domain wrappers invoke the same implementations with different x ranges
+- Monitor-first multi-GPU result semantics: `gather_fields=False` returns supported point/time and assembled plane/flux/mode payloads on `result_device` without a global field, while `gather_fields=True` preflights and gathers global electric last-step or DFT fields. y/z-normal `PlaneMonitor`/`FinitePlaneMonitor`/`FluxMonitor`/`ModeMonitor` payloads are tiled across owned x intervals; x-normal planes have one owner, with an explicit guard for ordinary `Ex` planes exactly on an internal split. Gathered `Result.save/load` and manifest-based `save_sharded/load_sharded` provide eager or lazy/gathered persistence without restoring live transport state
+- Qualification remains limited to two RTX A6000 GPUs connected by NV4 and a guarded real-valued forward subset: no three/four-GPU or PCIe-only qualification, Nsight trace, NCCL/multi-node execution, distributed adjoint, peer-aware CUDA Graph, advanced source families, x periodic/Bloch/symmetry, nonlinear/full-off-diagonal media, or SIBC claim is made. The final CUDA 13 acceptance observed 1.97385x strong speedup for the `257^3` vacuum case and exact single/two-GPU diagnostics, while the `129 x 65 x 65` case remained below break-even; see the [FDTD multi-GPU joint-solve guide](docs/reference/fdtd-multi-gpu-joint-solve.md) for the complete support matrix, persistence contract, test checkpoints, and benchmark table
 - Single- or multi-frequency DFT extraction through `frequency=` or `frequencies=[...]`
 - Source temporal frequency (`source_time.frequency`) remains distinct from simulation / monitor extraction frequencies; Maxwell does not infer extraction frequencies implicitly
 - ADE-based electric and magnetic dispersive-material updates for Debye, Drude, and Lorentz media

docs/README.md

@@ -0,0 +1,26 @@
+# Documentation
+
+The documentation tree is organized by document purpose. Solver and feature names stay in filenames rather than creating another directory layer.
+
+## Active documentation
+
+- [`plans/`](plans/): implementation plans and functional roadmaps that still have open work.
+- [`memory/`](memory/): current implementation notes and durable engineering context.
+- [`assessments/`](assessments/): current audits, gap analyses, and capability evaluations.
+- [`reference/`](reference/): maintained technical facts, capability matrices, and code-linked inventories, including the [FDTD multi-GPU joint-solve guide](reference/fdtd-multi-gpu-joint-solve.md).
+- [`releases/`](releases/): release notes.
+- [`validation/`](validation/): stable release-validation records.
+
+The active functional roadmap is indexed in [`plans/next-functional-2026-07/README.md`](plans/next-functional-2026-07/README.md).
+
+## Archive
+
+[`archive/`](archive/) contains completed or superseded plans, historical assessments, implementation snapshots, and design records. Archived material preserves engineering history but is not the current source of truth.
+
+## Placement rules
+
+1. Choose the directory from the document's purpose, not its solver backend.
+2. Use a descriptive filename such as `fdtd-cuda-graph-option-b.md`; do not add solver-specific directory layers.
+3. Move a document to `archive/` when its work is complete, its implementation model no longer exists, or a newer document supersedes its facts.
+4. Keep repository-wide user documents such as `README.md`, `CHANGELOG.md`, and `FEATURE_LIST.md` at the repository root.
+5. Keep executable development utilities under `scripts/`, never under `docs/`.

scripts/build_fdtd_cuda_prebuilt.py

@@ -43,7 +43,7 @@ def main() -> None:
     build_dir = Path(
         os.environ.get(
             "WITWIN_MAXWELL_FDTD_CUDA_BUILD_DIR",
-            Path(tempfile.gettempdir()) / "witwin_maxwell_fdtd_cuda_wheel" / "stable_abi_v1",
+            Path(tempfile.gettempdir()) / "witwin_maxwell_fdtd_cuda_wheel" / "stable_abi_v2",
         )
     )
     os.environ["WITWIN_MAXWELL_FDTD_CUDA_BUILD_DIR"] = str(build_dir)
@@ -61,6 +61,9 @@ def main() -> None:
             existing.unlink()
     target = cuda_build.prebuilt_extension_path()
     shutil.copy2(module_file, target)
+    marker = cuda_build.prebuilt_abi_marker_path(target)
+    marker.write_text(f"{cuda_build.STABLE_ABI_VERSION}\n", encoding="utf-8")
+
     print(f"Built prebuilt FDTD CUDA extension: {target}")