Add BH #2D hardware evidence harness and lightweight N-body home page

.github/workflows/ci.yml

@@ -38,6 +38,7 @@ jobs:
       - name: Check Barnes-Hut self-gravity reference
         run: |
           node tests/barnes-hut.mjs
+          node tests/nbody-view.mjs
           cargo test --manifest-path nbody/Cargo.toml --locked --offline
           cargo run --manifest-path nbody/Cargo.toml --locked --offline -- verify
           cargo run --manifest-path nbody/Cargo.toml --locked --offline -- verify-flat

.github/workflows/native-gpu.yml

@@ -3,9 +3,9 @@ name: Native GPU runtime
 on:
   push:
     branches: [main]
-    paths: ['runtime/**', 'rust/**', 'nbody/**', 'data/uff/**', 'tests/uff-reference.json', 'scripts/*gpu*', 'scripts/vast-runner.py', 'rust-toolchain.toml', '.github/workflows/native-gpu.yml', '.dockerignore', 'LICENSE', 'NOTICE.md']
+    paths: ['runtime/**', 'rust/**', 'nbody/**', 'data/uff/**', 'tests/uff-reference.json', 'tests/test_bh2d_hardware_sweep.py', 'scripts/*gpu*', 'scripts/bench-bh2d-hardware.py', 'scripts/bench-bh2d-hardware-launch.sh', 'scripts/bench-bh2d-hardware-launch.cmd', 'scripts/vast-runner.py', 'rust-toolchain.toml', '.github/workflows/native-gpu.yml', '.dockerignore', 'LICENSE', 'NOTICE.md']
   pull_request:
-    paths: ['runtime/**', 'rust/**', 'nbody/**', 'data/uff/**', 'tests/uff-reference.json', 'scripts/*gpu*', 'scripts/vast-runner.py', 'rust-toolchain.toml', '.github/workflows/native-gpu.yml', '.dockerignore', 'LICENSE', 'NOTICE.md']
+    paths: ['runtime/**', 'rust/**', 'nbody/**', 'data/uff/**', 'tests/uff-reference.json', 'tests/test_bh2d_hardware_sweep.py', 'scripts/*gpu*', 'scripts/bench-bh2d-hardware.py', 'scripts/bench-bh2d-hardware-launch.sh', 'scripts/bench-bh2d-hardware-launch.cmd', 'scripts/vast-runner.py', 'rust-toolchain.toml', '.github/workflows/native-gpu.yml', '.dockerignore', 'LICENSE', 'NOTICE.md']
   workflow_dispatch:
 
 permissions:
@@ -25,6 +25,12 @@ jobs:
           python-version: '3.12'
       - name: Check SSH runner and result extraction on Windows
         run: python runtime/tests/test_runner.py
+      - name: Check BH2D Windows launcher propagates Python failure
+        shell: cmd
+        run: |
+          call scripts\bench-bh2d-hardware-launch.cmd --particles 512 --dry-run
+          if %ERRORLEVEL% EQU 0 exit /b 1
+          exit /b 0
 
   runner-python:
     runs-on: ubuntu-24.04
@@ -60,6 +66,8 @@ jobs:
           bash -n scripts/run-gpu.sh scripts/bootstrap-gpu.sh
           python runtime/tests/test_runner.py
           python runtime/cuda/check_barnes_hut_layout.py
+          python tests/test_bh2d_hardware_sweep.py
+          sh scripts/bench-bh2d-hardware-launch.sh --dry-run --output "$RUNNER_TEMP/bh2d-plan" --particles 512,4096,8192
       - name: Install software Vulkan for compute validation
         run: |
           sudo apt-get update -qq

.github/workflows/pages.yml

@@ -29,6 +29,7 @@ jobs:
           node tests/physics.mjs
           node tests/app.mjs
           node tests/barnes-hut.mjs
+          node tests/nbody-view.mjs
           cargo test --manifest-path nbody/Cargo.toml --locked --offline
           cargo run --manifest-path nbody/Cargo.toml --locked --offline -- verify
           node scripts/build-site.mjs

NOTICE.md

@@ -9,14 +9,15 @@ and local PNG/WebM/JSON capture. Galaxy orbital equations and the accelerated
 renderer replace the original gate–centre–mouth transfer geometry.
 
 The adapted browser sources (`galaxy-core.js`, `app.js`, `renderer.js`,
-`index.html`, `style.css`, and the JavaScript smoke/application tests) retain
+`rotation-lab.html`, `style.css`, and the JavaScript smoke/application tests) retain
 **Mozilla Public License 2.0** file notices. The full licence is in
 [`LICENSES/MPL-2.0.txt`](LICENSES/MPL-2.0.txt). Their corresponding source is
 provided directly in this repository and static distribution.
 
-The new Rust sampler and native example, build scripts, generated Wasm module
-and browser payload, and benchmark are
-**Apache-2.0**, under the repository's existing [`LICENSE`](LICENSE).
+The new N-body browser entrypoint (`index.html`, `nbody-clock.js`, and
+`nbody-viz.js`), Rust sampler and native example, build scripts, generated Wasm
+module and browser payload, and benchmark are **Apache-2.0**, under the
+repository's existing [`LICENSE`](LICENSE).
 The existing repository licence has not been used to relicense the MPL files.
 
 VORTEX's reference photographs, artwork and historical stress-test reports are

README.md

@@ -7,14 +7,15 @@
 
 It began as an adaptation of the VORTEX 2.1.0 particle lab and now combines:
 
-- interactive rotation-curve visualization;
+- a lightweight default N-body simulator: 768 real interacting bodies, luminous sprites and short orbital trails;
+- interactive rotation-curve visualization in the preserved rotation-law instrument;
 - UFF, Newtonian, NFW, Burkert, MOND/RAR, and authored visual rotation laws;
 - deterministic Rust/WebAssembly sampling in the browser;
 - native Rust CPU execution with Float/libm and BAM32/Q2.30 LUT backends;
 - Vulkan/`wgpu` and NVIDIA CUDA compute paths;
 - memory-bounded exact-u64 logical addressing;
 - reproducible benchmark receipts, topology evidence, and archived scaling studies;
-- an opt-in Barnes–Hut resident self-gravity laboratory with a direct-force oracle and live quadtree visualization;
+- a Barnes–Hut resident self-gravity laboratory with a direct-force oracle and live quadtree visualization;
 - a GPU-oriented BH #2A Morton/Z-order + flat-cell CPU substrate with repeatable topology receipts;
 - a BH #2B1 explicit f32/u32 transfer ABI with executable Vulkan/WGSL flat-tree traversal and matched CUDA traversal source;
 - BH #2B2 multi-step resident self-gravity with persistent GPU state, GPU kick/drift/final-kick kernels, and an explicit CPU tree-rebuild boundary;
@@ -37,9 +38,10 @@ The v0.4.0 native-CPU evidence baseline remains archived at Zenodo as:
 
 | Layer | Current state |
 | --- | --- |
-| Browser instrument | Offline HTML/CSS/JS + bundled Rust/Wasm; no server or CDN required |
-| Browser logical population | Up to `2^32 = 4,294,967,296` logical stars on the Rust/Wasm path |
-| Browser rendered sample | Up to 65,536 stars per frame on Rust/Wasm + WebGL |
+| Browser N-body home page | 128–2,048 interacting bodies (768 default); planar Barnes–Hut gravity, leapfrog integration, glow/trails and orbit/zoom controls |
+| Rotation-law instrument | Preserved at `rotation-lab.html`; offline HTML/CSS/JS + bundled Rust/Wasm |
+| Rotation-law logical population | Up to `2^32 = 4,294,967,296` logical stars on the Rust/Wasm path |
+| Rotation-law rendered sample | Up to 65,536 stars per frame on Rust/Wasm + WebGL |
 | Native CPU runtime | Exact positive-u64 logical population, bounded resident sample up to 16,777,216 particles |
 | CPU projection backends | `float-libm` and `bam-lut-q30` |
 | Native GPU runtime | Rust/`wgpu`/Vulkan plus NVIDIA CUDA/CuPy RawKernel |
@@ -61,9 +63,13 @@ The earlier v0.4.0 archive remains the **before-state** for the CPU architecture
 
 ## 1. Browser instrument
 
-Open **`index.html`** directly in a modern browser. No install, local server, CDN, or network connection is required, including for the bundled Rust/WebAssembly engine.
+Open **`index.html`** directly in a modern browser for the N-body simulator. No install, local server, CDN, or network connection is required. The default 768 bodies all contribute mass; reducing the resident count reduces computation. Glow sprites and ten-step trails add visual density without adding simulated particles. The camera tilts a planar physical system; it is not a 3D force solver.
 
-The separate **`barnes-hut.html`** entrypoint is an opt-in resident self-gravity lab with a live quadtree overlay and direct-force error probes. It does not change the default rotation-law instrument.
+Choose a binary encounter, rotating disc or cold collapse. Pause, single-step, change the seed, inspect the tree, or pause for a direct-force audit. Physics uses a fixed time step on a 60 Hz schedule, independent of monitor refresh rate. Overload drops catch-up work rather than enlarging the time step or changing the body count. Reduced-motion preference starts paused and disables trails.
+
+The original prescribed-field instrument, including its Rust/WebAssembly sampling, UFF controls and exports, is preserved at **`rotation-lab.html`**. The detailed tree laboratory remains at **`barnes-hut.html`**. The following logical-population and sampling controls describe the rotation-law instrument.
+
+The separate **`barnes-hut.html`** entrypoint is an opt-in resident self-gravity lab with a live quadtree overlay and direct-force error probes. The default browser entrypoint is the N-body simulator at **`index.html`**, while the preserved rotation-law instrument remains at **`rotation-lab.html`**.
 
 The browser instrument lets you change morphology, mass model, viewing geometry, and time while watching the galaxy and its rotation curve respond together.
 
@@ -272,6 +278,16 @@ BH #2D is the active optimization path. It retains BH #2C and BH #2B2 as executa
 
 BH #2D correctness is exercised through software Vulkan in CI, but **software-Vulkan timings are not hardware performance evidence**. A hardware speed claim requires a non-software adapter receipt.
 
+The remaining BH #2D evidence item has a fail-closed real-GPU sweep runner:
+
+```bash
+sh scripts/bench-bh2d-hardware-launch.sh \
+  --output runs/bh2d-hardware-sweep \
+  --adapter 0
+```
+
+It runs a source-pinned 512→65,536 resident-body sweep by default, requires hardware receipts, validates the frozen correctness/oracle gates, and hashes each receipt/log into a scaling manifest. The manifest is evidence for that exact source/workload/adapter, not an automatic production-promotion claim.
+
 `--allow-software` enables verification through software Vulkan and must not be interpreted as hardware GPU performance evidence.
 
 The CUDA checker freezes ABI/source parity without claiming CUDA execution where no NVIDIA CUDA run occurred.
@@ -467,6 +483,9 @@ sh scripts/test-cpu-runtime.sh
 # Retro CPU portability
 sh scripts/test-retro-cpu.sh
 
+# GPU evidence-runner host-side tests
+python3 tests/test_bh2d_hardware_sweep.py
+
 # CUDA host-side tests
 python3 tests/test_cuda_u64.py
 ```

ROADMAP.md

@@ -199,6 +199,8 @@ Stable radix ordering starts from resident-body order and preserves that order f
 
 The initial CI fixture proves correctness through Mesa Vulkan but does not establish a hardware speedup. Real GPU performance evidence remains an explicit BH #2D completion item.
 
+The repository now includes `scripts/bench-bh2d-hardware.py` to capture that evidence without weakening the boundary. The runner requires a clean tracked source tree and `--require-hardware`, validates the BH #2D correctness/determinism gates at every point, preserves repeat-matched BH #2C comparison through 4,096 bodies, and emits a source-pinned scaling manifest with receipt/log hashes. The harness is implementation support; BH #2D remains evidence-pending until a real-GPU manifest exists.
+
 ## BH #2E — Hardware Promotion and Scaling
 
 After BH #2D obtains real-hardware receipts, the next promotion rung is:
@@ -397,3 +399,7 @@ All of the following remain useful research directions, but they are explicitly
 These items were deferred while PE #15 was active. v0.6.0 closed that phase; any future work now requires an explicit phase with its own evidence boundary.
 
 The previous detailed designs are preserved by the immutable `v0.5.0` source archive and tag; they are not discarded, only postponed.
+
+## Browser N-body observatory
+
+The home page now runs the existing planar Barnes–Hut solver with a modest 768-body default, selectable 128–2,048 bodies, fixed-rate integration, glow/trails and camera controls. The original prescribed-field visual is preserved at `rotation-lab.html`; `barnes-hut.html` retains the detailed tree lab. This UI upgrade supplies no native GPU timing evidence and does not change BH #2D evidence-pending or BH #2E promotion gates.

barnes-hut.css

@@ -69,3 +69,13 @@ footer { min-height: 40px; border-top: 1px solid var(--line); color: var(--muted
   .readouts div { border-right: 0; border-bottom: 1px solid var(--line); }
   .readouts div:last-child { border-bottom: 0; }
 }
+
+/* N-body observatory controls. */
+.gesture-note { color: var(--muted); font-size: 0.7rem; align-self: center; margin-left: auto; }
+#nbodyCanvas { touch-action: none; cursor: grab; }
+#nbodyCanvas:active { cursor: grabbing; }
+:focus-visible { outline: 2px solid var(--hot); outline-offset: 3px; }
+input[type="range"], input[type="checkbox"] { accent-color: var(--hot); }
+@media (max-width: 560px) {
+  .gesture-note { margin-left: 0; width: 100%; line-height: 1.5; }
+}

barnes-hut.html

@@ -18,7 +18,7 @@
       <p class="eyebrow">QSOL-IMC / GALAXY / SELF-GRAVITY LAB</p>
       <h1>BARNES–HUT</h1>
     </div>
-    <nav><a href="index.html">Main instrument</a><a href="https://github.com/QSOLKCB/GALAXY">Source ↗</a></nav>
+    <nav><a href="index.html">N-body simulator</a><a href="https://github.com/QSOLKCB/GALAXY">Source ↗</a></nav>
   </header>
 
   <main class="layout">
@@ -48,7 +48,7 @@ <h1>BARNES–HUT</h1>
       <details class="note" open>
         <summary>What changed?</summary>
         <p>This lab is a separate self-gravitating execution path. Every resident body contributes mass to a deterministic quadtree. Distant cells may be replaced by their centre of mass when <code>s / d &lt; θ</code>; nearby cells are opened. A kick–drift–kick leapfrog step rebuilds the tree after the drift.</p>
-        <p>The main GALAXY rotation-law instrument remains unchanged. Its huge logical populations are sampled independent test particles; they are not silently reinterpreted as mutually interacting bodies. This lab keeps the resident population explicit because self-gravity creates cross-particle coupling.</p>
+        <p>The <a href="rotation-lab.html">GALAXY rotation-law instrument</a> remains available separately. Its huge logical populations are sampled independent test particles; they are not silently reinterpreted as mutually interacting bodies. This lab keeps the resident population explicit because self-gravity creates cross-particle coupling.</p>
       </details>
     </section>
 

docs/BARNES-HUT.md

@@ -2,7 +2,7 @@
 
 GALAXY now has two deliberately distinct dynamics families:
 
-1. **Prescribed-field / test-particle modes** — the existing browser, CPU and GPU paths. These retain huge logical address spaces because individual resident particles do not affect one another.
+1. **Prescribed-field / test-particle modes** — the rotation-law browser instrument (`rotation-lab.html`), CPU and GPU paths. These retain huge logical address spaces because individual resident particles do not affect one another.
 2. **Resident self-gravity** — the Barnes–Hut laboratory. Every resident body's mass contributes to the force field, so the complete interacting resident set is explicit and cannot be substituted by independent logical-u64 tiles.
 
 ## Why Barnes–Hut
@@ -55,6 +55,9 @@ A future GPU implementation should preserve this CPU/direct reference as its cor
 
 ## Visualization
 
+`index.html` is the default lightweight N-body observatory. It reuses the verified browser force solver with 768 resident bodies by default (128–2,048 selectable), fixed-rate leapfrog scheduling, seeded presets, optional trails and luminous sprites. Rendering contributes no additional gravitational mass. Orbit/zoom controls project the planar dynamics without modifying them. At 4× speed on an ordinary 60 Hz display, each rendered frame advances four physics steps. After a render gap the scheduler may batch up to 24 steps, derived from the 100 ms elapsed-time cap, then discards any remaining overload debt; hidden tabs suspend physics and under sustained load simulated time advances more slowly. Force probes pause the simulation so their result remains attached to the displayed state.
+
+
 `barnes-hut.html` is an offline browser laboratory. It renders the resident bodies and can overlay quadtree cells while the system evolves. Controls expose:
 
 - resident body count;
@@ -289,6 +292,34 @@ Kernel families are batched into command buffers to avoid measuring one CPU/GPU
 
 Receipts classify the selected adapter as either `software-validation` or `hardware`. `--require-hardware` rejects software adapters. Mesa/llvmpipe timing is therefore correctness and diagnostic evidence only; a hardware performance claim requires a receipt from a real GPU.
 
+### Real-hardware sweep
+
+`scripts/bench-bh2d-hardware.py` is the evidence runner for the remaining BH #2D hardware item. It executes a strictly increasing resident-body sweep through `galaxy-bh-gpu-tree-parallel --require-hardware`, refuses a dirty tracked source tree or an existing evidence directory, validates every receipt, requires one adapter identity for the complete sweep, and hashes each receipt/log into `manifest.json`.
+
+The default sweep is:
+
+```text
+512 -> 1024 -> 2048 -> 4096 -> 8192 -> 16384 -> 32768 -> 65536
+```
+
+Counts through 4,096 retain repeat-matched BH #2C timing. Larger points explicitly require the BH #2C and bounded full-oracle skips already encoded by the verifier; they do not silently drop those gates.
+
+Example on a real GPU:
+
+```bash
+sh scripts/bench-bh2d-hardware-launch.sh \
+  --output runs/bh2d-hardware-sweep \
+  --adapter 0
+```
+
+A completed scaling manifest is hardware evidence for the exact recorded source, workload and adapter. It does **not** by itself promote BH #2D to production or establish a host-independent speed claim.
+
 ## Next rung
 
 After real-hardware BH #2D receipts exist, BH #2E can evaluate scaling, memory/capacity limits and production promotion while retaining BH #2A/B2B2/B2C as frozen oracles.
+
+### Hardware harness launch and toolchain boundary
+
+Launch real evidence capture through the clean launcher: `sh scripts/bench-bh2d-hardware-launch.sh ...` on POSIX or `scripts\\bench-bh2d-hardware-launch.cmd ...` on native Windows. The launcher starts isolated Python under a newly constructed environment so inherited loader-injection variables are not resident in the evidence process. Direct Python execution is rejected. Imported use by the unit-test runner is for host validation only.
+
+The harness now selects and records a platform-specific tool boundary. Linux uses the system compiler/linker/Git path, macOS uses the native system paths plus standard Homebrew locations, and Windows records Git plus any active MSVC/LLVM linker tools before narrowing the build PATH to those resolved tool directories. Cargo and rustc are recorded separately from rustup proxies and invoked through their resolved binaries. The OS, installed system libraries and toolchain are trusted host prerequisites; this is provenance checking, not a sandbox against a hostile host. Git checks clear ambient `GIT_*` selectors and bind the checkout explicitly, including linked worktrees. Relative `CARGO_HOME` is interpreted from Cargo's repository working directory. Logs preserve raw bytes, receipt parser failures enter the failed-manifest path, and workload values are validated before any output is created. Manifests serialize strict JSON.

docs/ENGINE.md

@@ -95,7 +95,7 @@ represent a 2³² logical population. WebGL context loss stops the current recor
 and switches to a fresh Canvas element with the smaller sample.
 
 The payload is base64 in an external classic script, avoiding fetch and file-URL
-CORS restrictions when opening `index.html` directly. It needs no `eval`, no
+CORS restrictions when opening `rotation-lab.html` directly. It needs no `eval`, no
 worker, no remote package, and no cross-origin-isolation headers. CSP explicitly
 permits Wasm compilation using `wasm-unsafe-eval`. Compilation failure uses the
 bounded JavaScript fallback and is reported in the interface.