Milestones are marked COMPLETE only after the code has been built, run, and its behaviour checked. For rendering work that means a frame was captured offscreen and looked at; for physics it means a test asserts the numbers.

Current state: all twelve success criteria met, plus a visual polish pass and a UI hardening pass. 160 unit tests and a 41000-check application self-test, all passing (ctest --test-dir build).

Status: COMPLETE

Implemented: CMake/Ninja project, GLFW 3.4 + GLM 1.0.1 + Dear ImGui v1.91.5 via FetchContent, vendored glad2 loader; Window, Clock, Input, Application main loop, asset resolution, and a hand-written PNG encoder for offscreen capture.

Build: cmake -S . -B build -G Ninja && cmake --build build, clean.

Run: build/bin/universe-sim.exe --screenshot docs/images/m1_window.png --frame 5

Tests: 6 passing (tests/test_units.cpp, unit formatting).

Observed: context reports NVIDIA GeForce RTX 4090 | 3.3.0 | GLSL 3.30. docs/images/m1_window.png is a uniform #050508 field at the requested size, matching the clear colour -- which also confirms the PNG encoder emits valid files.

Problems found and fixed:

1. POST_BUILD asset staging was impossible here. CMake wraps custom commands in a nested cmd /C "cd /D <path> && ...", and this repository sits under a directory named AI & Media; cmd re-parsed the ampersand as a command separator and the copy failed with 'Media' is not recognized. Replaced with runtime asset resolution, which also allows shader hot-reload.
2. glfwGetKey spammed GLFW_INVALID_ENUM ~160 times per frame: the input poll ran to key code 511 but GLFW_KEY_LAST is 348.
3. Unit formatter chose megametres for Earth's radius; astronomy uses km.

Known issues: none.

Status: COMPLETE

Implemented: triangle-fan circle mesh from centre + rim vertices; CelestialBody with integrate via the shared integrators; fixed-timestep accumulator; axis-aligned bounds with configurable restitution and friction.

Run: build/bin/universe-sim.exe --scene bounce

Tests: 9 (tests/test_kinematics.cpp). Free fall matches y = -1/2 g t^2 to 1e-12; fall time from 10 m matches sqrt(2h/g); impact speed matches sqrt(2gh); a perfectly elastic bounce returns to the drop height; restitution 0.5 returns to a quarter of it (the e^2 law); friction halves tangential speed at f = 0.5; the ball never leaves the box over 20 000 steps; and 120 Hz and 480 Hz stepping agree to 1e-12, which is the property that makes physics frame-rate independent.

Observed: frames captured at 0.4 s, 1.5 s and 2.6 s of simulated time show barely-moved balls, then visibly parabolic trails, then the V of a completed bounce.

Problems found and fixed:

1. The test framework's CHECK_REL divided by |expected|, so every assertion with a negative expected value -- potential energy, inward acceleration -- reported a perfect match as -1 and failed. Six tests were failing for this reason alone.
2. The momentum conservation test normalised against a total that cancelled to zero, so it was measuring nothing.

Status: COMPLETE

Implemented: O(N^2) pairwise gravity, each unordered pair visited once with equal-and-opposite accelerations applied together; three configurable singularity strategies (none, Plummer softening, minimum distance).

Tests: 10 across test_gravity.cpp (vector maths, acceleration and stabilisation). Acceleration matches the closed form for two 1 kg masses 1 m apart; doubling the distance quarters the acceleration exactly; m*a sums to zero across an unequal-mass pair to 1e-14; symmetric masses cancel at the midpoint; softening matches the analytic Plummer value to 1e-10 and stays finite at zero separation; a fixed body never accelerates.

Observed: the attract preset shows three masses falling together from rest, each sitting in its own visible well.

Status: COMPLETE

Implemented: explicit Euler, symplectic Euler, velocity Verlet, RK4, switchable at runtime; energy/momentum/angular-momentum diagnostics with drift tracking; trails with a configurable sample interval and reference frame.

Tests: 13 (tests/test_integrators.cpp). Velocity Verlet is exact for a constant field; RK4 tracks cos(t) to 1e-8 over a full period at 200 steps; explicit Euler grows an oscillator's energy from 0.5 past 10 in 10 000 steps while symplectic Euler stays in 0.45-0.55; halving the step cuts Verlet's position error by 3-5x, confirming second order; a circular orbit returns to its start after one period to 1e-4 of its radius; momentum is conserved to 4e-14 over 20 000 steps; identical initial conditions reproduce bitwise.

The energy test asserts boundedness, not smallness: the peak drift over orbits 1-25 is compared with orbits 26-50, which is the property that actually distinguishes a symplectic integrator.

Observed: Earth completes exactly one revolution per Julian year (within 0.02 radians) and its orbital radius stays within 0.5% of 1 AU.

Status: COMPLETE

Implemented: procedural UV spheres with pole degeneracy handled; Shader, Mesh, MeshFactory, Renderer; fly and orbit cameras with mouse-look, WASD, vertical movement, scroll speed control and focus-on-body; camera-relative (floating origin) rendering; lit bodies with emissive stars, rim light and a subtle specular.

Observed: verified across all fourteen presets; the inner system shows lit spheres, elliptical trails and correct relative sizes.

Problems found and fixed:

1. Uniform radius exaggeration is unusable at solar-system scale. The Sun is 109 Earth radii, so the multiplier that made Earth visible drew the Sun wider than Earth's whole orbit -- visible immediately in the first capture. Replaced with a power law gain * r^exponent, with each scene solving its own gain from a named reference body.
2. The gain was solved from the largest body in the scene, so the Earth-Moon preset -- which still contains the Sun, 2 992 units off screen -- drew the Earth six pixels across.

Status: COMPLETE

Implemented: metresPerUnit / drawn-radius exaggeration / physical radius kept strictly separate; fourteen presets from real IAU/NASA values; JSON scene configuration in configs/; timeScale from 0.01x to 1e9x with substepping and a step budget; pause, single-step, reset.

Tests: 29 -- 15 in test_scenes.cpp and 14 in test_json.cpp. Every preset is well formed with unique keys and finite state; astronomical presets have zero net momentum; the inner system's orbital radii stay within 5% over ten simulated years; the Moon stays between 3.0e8 and 4.6e8 m of Earth over two years; energy drift over a decade peaks below 1e-5.

Config round-trips are asserted bitwise: the inner system is run 4000 steps from the built-in preset and from its saved-and-reloaded copy, and the summed positions must be exactly equal.

Status: COMPLETE

Implemented: wireframe or shaded XZ patch, displaced per vertex on the GPU by the softened Newtonian potential of up to 24 nearby bodies, smoothly saturated; configurable strength, opacity, resolution, extent, depth and well softening. Documented throughout as a visualisation.

Observed: single well under the Sun, a merged double well for the binary, three overlapping wells for the three-body scene.

Problems found and fixed:

1. The patch was centred under the camera rather than under what the camera was looking at. At the binary preset's 70 unit camera distance over a 55 unit patch, the stars sat off the edge of their own sheet and had no wells at all.
2. The distance fade was keyed to the grid extent, which erased the sheet entirely whenever the camera was farther out than 1.6x that -- the full solar system view had no grid whatsoever. It now fades against the distance to the patch's far corner.
3. Far-off bodies flattened everything. The Sun, 2 992 units from a 26 unit patch in the Earth-Moon preset, dominated the normalisation and pushed the whole sheet into saturation. Wells are now culled beyond three grid extents and normalised against the heaviest body actually on the patch.
4. Wells were too narrow to read at the default softening; widened from 0.02 to 0.055 of the extent.

Status: COMPLETE

Implemented: menu bar with live FPS, simulated time and body count; Scenes, Simulation, System diagnostics, Rendering, Bodies + inspector, Spawn, and Help panels; screen-space body labels; click-to-select via ray picking; Tab cycling; focus, delete, duplicate; eight spawn presets; spawn ahead of the camera or at exact coordinates.

Tests: 16 view-maths tests (test_viewmath.cpp), made possible by moving picking and scale maths into sim/ViewMath so they need no GL context. Picking hits the body under the ray, ignores bodies behind the camera, prefers the nearest of two, uses the drawn radius rather than the physical one, and works at astronomical coordinates.

Problems found and fixed:

1. Panels used hard-coded pixel positions and ran off screen at 1280x720; now laid out against the live viewport.
2. Widget labels were clipped. ImGui puts a label to the right of a widget and defaults the widget to full width, so labels ran past the panel edge.

Status: COMPLETE

Implemented: Schwarzschild radius computed and displayed for every body and drawable at true scale; compact-object presets with explicit "this is not a black hole" wording in the UI and docs; collision modes (ignore / elastic / merge) with volume-additive radii and momentum conservation; --spawn and --settle command-line options so runtime insertion can be verified from a script through the same code path the UI uses.

Tests: 11 collision tests (test_collisions.cpp) plus the compact-object stability tests in test_scenes.cpp. A merge conserves mass and momentum, places the result at the centre of mass, and adds volumes rather than radii; an elastic collision at restitution 1 conserves both momentum and kinetic energy to 1e-12.

Observed: a neutron-star-like mass dropped into the inner system produces dramatic, entirely finite orbital disruption -- Mercury's distance grows 2329%, Earth's 815%, with no NaN, no infinity and nothing exceeding c.

Finding worth recording: a close encounter is finite but not accurate. Dropping a compact object through a star moves the total energy by a factor of thousands. That is timestep resolution, not the force law, and the two are distinguishable by refining the step -- which a test now does at 600 s, 150 s and 37.5 s, requiring the drift to fall monotonically and by more than 10x overall. Documented in docs/PHYSICS.md.

Status: COMPLETE

Implemented: HDR render target (multisampled RGBA16F) with a bright pass, half-resolution separable Gaussian bloom, ACES filmic tone mapping, exposure and a subtle vignette; a procedural fixed-seed starfield drawn as round point sprites with a power-law brightness distribution and blackbody-ish tints; a graceful fallback to direct rendering when the framebuffer cannot be created.

Emissive bodies now output well above 1.0 on purpose, so the bright pass finds them and the tone map rolls their cores to white while the bloom halo keeps the star's colour. That is the whole reason the target has to be floating point: at 8 bits the overflow is clipped before bloom can see it.

Run: --no-post and --no-stars disable the new work, which is how the fallback path is exercised rather than assumed.

Observed: captured before and after across the presets. The Sun gains a real halo, the previously empty black void has a sky, and the grid reads far better against it. UI panels remain crisp because ImGui draws after the composite -- confirmed in docs/images/hero.png.

Tuning: the first pass had bloom washing out the inner planets and the starfield competing with the grid, so the threshold went 1.0 -> 1.15, intensity 0.75 -> 0.55, star brightness 1.0 -> 0.75 and grid opacity 0.42 -> 0.34.

Known issue: stars are visible through the grid sheet, including "below" it. That is correct for a transparent visualisation plane rather than a floor, but it does read oddly at shallow camera angles.

Status: COMPLETE

Implemented: reset-to-defaults for rendering, simulation and camera settings, individually and together, from both a Reset menu and per-panel buttons; a barycentre-following camera mode; merge counts surfaced in the diagnostics panel; a deterministic frame-sequence mode (--sequence) used to render the demo clips; and --selftest, which drives every UI-reachable state transition against a real GL context.

Bugs found by reading and by the self-test, not by using the app:

1. Use-after-free in Duplicate. spawnBody() push_backs into the body vector, which can reallocate. The handler then read body->name through a pointer into that vector to build its status message.
2. Dangling references after deletion. Removing a body left the selection, the camera focus and -- worst -- the trail reference frame pointing at an id that no longer existed. A dead trail frame silently reinterprets every stored sample as inertial, so all trails jump. Deletion now goes through Application::deleteBody, which repairs all three, and merges use the same path.
3. The camera never moved in sequence mode. runSequence set the orbit target and swept the yaw but never called Camera::update, so the camera stayed where it started and the sweep merely panned. Scenes whose bodies move drifted out of frame, which is why the first demo clips were mostly empty space.
4. The three-body preset was using the wrong speed. It set 0.9 * sqrt(GM/r), the circular speed about a single central mass, for a configuration that has three. The correct balance for an equilateral triangle of equal masses is v = sqrt(G m / (sqrt(3) r)); with the wrong value the triangle tore itself apart in under two years instead of orbiting. Now pinned by a test.
5. Unreadable time-scale buttons. "%.0g" rendered 1e4 as 1e+04x, wider than the button it sat in.

Self-test: 11 scenes x (load, step, four resets, spawn, focus, delete, delete the trail reference, all four integrators, all three stabilisation modes, delete every body, reset while empty, reload) plus picking and JSON round-trips. 41005 checks, 0 failures.

One finding kept rather than fixed: in the inner-system view the Moon sits inside the Earth's drawn sphere and cannot be clicked. That is a consequence of the radius exaggeration, not a picking bug, so the self-test asserts the weaker correct property -- a missed pick must be explained by an enclosing body -- and reports the count.

• Any relativistic physics. The seam is GravitySystem::accelerations plus the ForceModel interface; see the end of docs/PHYSICS.md.
• Barnes-Hut, octrees or compute-shader gravity. O(N^2) is entirely adequate for the tens of bodies these scenes use.
• Bloom and other post-processing.
• Real ephemeris data. Orbits start as circles at the semi-major axis, which keeps initial conditions deterministic and legible.