Two issues from code review: (1) the noise coordinate used phi=atan2
which has a branch-cut discontinuity at ±π, producing a visible radial
seam. Switch to Cartesian pos.x/pos.z (continuous, seam-free). (2)
value_noise3 summed three hash components per corner via dot(.,(1,1,1)),
giving a [-1.5,1.5) range instead of [0,1). Use a single scalar hash
(.x) per corner and drop the +0.5 recenter.
The disk's noise was two sin() terms (phi*8, phi*23) producing geometric
bands. Replace with domain-warped FBM: a 3-octave value-noise field warps
the sampling coordinate of a 4-octave FBM, giving the feathered/smoky gas
texture of the Gargantua reference. The Keplerian shear (rot ∝ 1/r^1.5)
is retained as the flow time term so inner radii rotate faster.
Switch the offscreen target from Bgra8UnormSrgb to Rgba16Float for HDR
headroom (disk brightness can exceed 1.0 without clamping). Set offscreen
camera order to -20 (the bloom pipeline, added next, needs cameras ordered
so offscreen renders first). Introduce three marker components:
- Nudgable: all render cameras carry it; nudge_camera now queries this
instead of a hardcoded Or<(OffscreenCamera, UpscaleCamera)>.
- QuadScaleFactor(fx, fy): fraction of offscreen res each quad fills;
resize_offscreen rescales each quad independently via this.
- CompositeQuad: the final quad renders to the window (rescaled against
window res, not offscreen).
The upscale path still works (textureSample reads float textures);
ACES tone-mapping is added in a later commit.
10-task plan (Tasks 0-9) for the HDR bloom + FBM disk + ACES + AA
pipeline. Each task is a standalone commit: Rgba16Float offscreen →
FBM disk → gaussian-speck stars → brightpass → blur pyramid →
composite+ACES → BloomQuality enum → Quality panel → verification.
Conventions established upfront (camera orders, RenderLayers, the
QuadScaleFactor/CompositeQuad/Nudgable markers) so each task references
them rather than re-deriving. The naga var<private> array fix for the
blur kernel is baked into the shader from the start.
Design for raising the renderer from geometrically-correct-but-flat to
Gargantua-reference fidelity. Four user-selected traits:
- HDR color + ACES tone mapping (Rgba16Float offscreen, Narkowicz fit)
- Multi-pass HDR bloom (bright-pass → down-pyramid → up-pyramid → composite)
- Domain-warped FBM disk texture (replaces two-sine noise)
- Anti-aliasing via render_scale + gaussian-speck stars
All quality options configurable via a dedicated egui Quality panel,
with tiered web defaults (Low bloom, no MSAA/star-AA) vs desktop (High).
BloomQuality::Off falls back to the original LDR path for WebGPU
browsers without float-filterable support.
Geodesic integrator (deriv/rk45_step/is_captured_rk45) untouched —
the physics.rs mirror contract is unaffected. No new tests.
At 1.318 rad (~75.5 deg) the disk is nearly vertical: its normal points
almost along +Z while the camera yaw axis is world-Y, so the two axes
sit ~75 deg apart. Orbiting via yaw (a horizontal circle around world-Y)
therefore shears the disk through view -- full circle, squashed ellipse,
edge-on sliver, back circle -- which reads as the disk swinging or the
scene rolling.
0.45 rad (~25.8 deg) brings the disk near-horizontal (normal approx
(0, 0.90, 0.43)), shrinking the yaw-axis / disk-normal misalignment
from ~75 deg to ~26 deg. The swing is correspondingly muted; orbiting
feels like turning around the hole rather than shearing across a wall.
Both defaults are updated in lockstep: BlackHoleParams (the live runtime
default) and BlackHoleUniforms (the bootstrap uniform filled before the
first mirror_params frame).
basis() derived `right` from `forward × world_up`, which degenerated to
a zero vector when forward aligned with world-Y (pitch = ±π/2) — the
cross product vanished, normalization produced garbage, and crossing
the pole flipped the image 180°.
Right now depends on yaw alone: Vec3::new(cos yaw, 0, -sin yaw). It is
a unit vector in the y=0 plane for any yaw, so it never goes to zero
regardless of pitch. In the original upper hemisphere (cp > 0) this is
exactly the old value after normalization, so existing views are
unchanged; it additionally stays perfectly horizontal for all pitch,
fixing a subtle roll the cross-product form introduced below the equator.
With the singularity gone, pitch is unclamped from [0.05, π-0.05] to
[-π+0.05, π-0.05]: the camera can now cross both poles and reach below
the disk plane without flipping. The ±0.05 margin only avoids the exact
origin-axis points where forward = ∓Y and yaw becomes meaningless.
yaw is also wrapped to (-π, π] on each drag so it cannot drift out of
the UI slider range (which would snap the view when the slider is later
touched); cos/sin periodicity makes the wrap visually lossless.
The UI pitch slider range is widened to match. Stale "gimbal pole"
comments that described the old restriction are corrected.
basis() still returns (eye, forward, right, up), so the uniform bridge
in plugin.rs and both shaders are untouched.
The docs described Phase 2 as 'future work' while the code had already
shipped it: spin is wired through the uniform into the shader's Kerr deriv(),
the integrator is adaptive RK45, disk_inner is ISCO-derived, render_scale
runs a real offscreen render-to-texture + upscale pass, and the UI exposes a
Spin slider with live ISCO/Horizon readouts. Anyone reading the docs got a
wrong picture of completeness.
README:
- Intro: 'Schwarzschild' -> 'black-hole (Kerr at spin>0, Schwarzschild at 0)'.
- How it works: RK4 -> adaptive Dormand-Prince RK45; add the frame-dragging
term to the bending accel; capture radius r+(chi); ISCO-tracked inner edge.
- Controls: add Black Hole section (Spin, ISCO/Horizon); render_scale slider.
- Performance: render_scale now wired (0.75 desktop / 0.5 web), not reserved.
- Status: Phase 1 shipped; Phase 2 implemented-pending-human-validation;
Phase 3 (exact pseudo-Hamiltonian) is the actual future work.
AGENTS.md:
- Intro + mirror contract: name the Kerr/RK45 functions the mirror must track
(kerr_bending_accel, rk45_step, is_captured_rk45) and the exact shader lines.
- Conventions: 'render_scale and spin are reserved' is false now — both wired;
update the perf-lever and web-defaults notes accordingly.
- Reference docs: list both Phase 2 docs, not just Phase 1.
Phase 2 plan (docs/superpowers/plans/2026-07-13-...-kerr.md):
- Status banner: Tasks 1-7 implemented+committed, Task 8 (human visual/perf
checklist) is what remains.
- Check the Task 1-7 step/commit boxes; leave all Task 8 manual steps and the
visual/fps acceptance items unchecked. Check only the code-verifiable
acceptance items (cargo test green, RK45+render_scale running, Spin slider
shipped).
Phase 2 shipped the Kerr deriv() and the adaptive RK45 integration loop in
the shader (black_hole.wgsl:260, 320-390), but the CPU mirror in physics.rs
only covered the single-step Kerr derivative (kerr_bending_accel), not the
loop. That left the AGENTS.md 'CPU <-> shader mirror' contract broken for
Phase 2: the spin-dependent capture radius, the adaptive step control, and
the budget = accepted-steps-only / dt_min forced-accept semantics existed
only on the GPU, with nothing testable on the CPU side.
Add a faithful CPU mirror and the loop-level tests that close the gap:
- rk45_step(): Dormand-Prince step mirroring black_hole.wgsl:260-285, using
kerr_bending_accel so it is also the Phase 1 step at chi=0.
- is_captured_rk45(): the full adaptive loop mirroring black_hole.wgsl:320-390
— same seeding (total_path / steps), same reject/retry at dt_min (the
forced-accept floor that prevents infinite retry), same accept-then-refine,
and the spin-dependent capture radius r+(chi) via kerr_horizon.
Tests (tests/physics_test.rs):
- spin=0 RK45 loop captures below bcrit and escapes above (degeneracy).
- spin=0.9 still captures a b~2.0 ray (horizon shrinks but not past b<bcrit).
- capture set does not grow as spin increases across a bcrit-straddling sweep.
- rk45_step error shrinks monotonically with dt (the property the adaptive
loop's reject/accept decision depends on).
Note: probing the error scaling showed it falls between 2nd and 4th order in
dt, not a clean 5th, because of the per-stage normalize() projection the
shader (and now the mirror) applies. The loop only relies on monotonicity, so
this is correct-as-shipped; documented in a doc-comment on rk45_step so a
future reader is not misled by the '45' in the name.
cargo test: 17 passed (3 Phase 1 inline + 14 integration, up from 12).
Fixes the following clippy warnings in plugin.rs:
1. clippy::too_many_arguments: Suppressed on spawn_fullscreen_quad since Bevy system functions naturally take many arguments as resources/queries.
2. clippy::field_reassign_with_default: Refactored instantiation of BlackHoleMaterial to set fields in the struct literal instead of post-creation assignment.
3. clippy::type_complexity: Suppressed on resize_offscreen and nudge_camera systems as Bevy queries with Filters/Disjunctions are inherently complex.
Clippy warned that the manual Default implementation for BlackHoleMaterial
could be derived (clippy::derivable_impls). We now derive Default directly
on the struct and remove the manual implementation block.
Fixes clippy::approx_constant errors caused by using the hardcoded value 3.14
for camera yaw and accretion disk tilt sliders. Using std::f32::consts::PI
resolves the errors and uses the precise constant value.
Clippy warned about excessive precision in the BCRIT float literal (2.5980762).
Since f32 has 24 bits of mantissa (approx 7 decimal digits of precision), the
excessive decimal digits are truncated to 2.598_076 to compile cleanly under
clippy::excessive_precision.
The integration test still passes as the difference is well within the 1e-5 limit.
resize_offscreen declared two Query<&mut Transform> params (FullscreenQuad
and UpscaleQuad filters). Bevy's conflict checker does not treat With<T>
filters as disjoint access — both queries declare write access to the same
component, so the app panicked at startup with error[B0001].
Merge the two queries into a ParamSet (borrowed one at a time) and move
them to the end of the parameter list per the Params convention.
- Set default pitch to 0.7 rad to keep the camera away from the π/2 gimbal pole, preventing horizontal drag from degenerating into an in-plane roll.
- Adjust scroll zoom logic to divide instead of multiply by the zoom factor, so that scrolling up zooms in (reduces distance) rather than out.
Compact onboarding file capturing hard-won, non-obvious repo facts:
- Dual-target build (desktop + web/WebGPU) and the wasm-only
web_sys_unstable_apis rustflag (.cargo/config.toml) that is a no-op
on desktop — easy to break during cleanup.
- Always run --release for visual checks; the debug ray tracer is too
slow and will look broken.
- The CPU<->shader mirror in src/physics.rs: it is a hand-maintained
duplicate of black_hole.wgsl kept solely so the capture/escape
boundary is testable. Editing one side without the other makes tests
pass on code the shader contradicts.
- The three Bevy 0.19 gotchas that cause the recurring grey screen
(nudge_camera for issue #24448, EguiPrimaryContextPass placement,
real storage buffer asset instead of default handle) — exactly what
this branch's recent commits fix.
- Reserved-but-intentionally-unwired fields (render_scale, spin) and
the Rs=1 natural-units convention so an agent does not 'fix'
intentional dead code or invent unit conversions.
- Git workflow: self-judge whether to commit, granular detailed
commits, never push.
No code changes.
The app showed only the camera clear color (grey srgb 43,44,47) — the
fullscreen quad's fragment shader produced no output. Two issues; the
second was the real blocker.
1. render/material.rs: skybox.wgsl declared binding 1 as texture_cube<f32>
but the material's #[texture(1)] defaulted to D2. The mismatched
bind-group layout vs. shader caused the pipeline to fail to specialize.
Fix: dimension = "cube" on the texture attribute so the layout matches.
2. SHADER COMPOSITION (the actual blocker): the shader was split across
naga_oil modules (ray_gen, geodesic, stars, disk, planets, grid, skybox)
pulled in via #import singularity::... This compiled and validated with
zero errors, but at runtime calling ANY cross-module-imported function
made the fragment output nothing (only the clear color showed). Local
functions were fine. Confirmed by bisecting a minimal repro with the
user watching the screen: local fn -> renders; imported fn (even pure
math) -> grey.
Rather than chase the naga_oil 0.22 / Bevy 0.19 composition bug, inline
every function into black_hole.wgsl and drop the #import singularity::*
lines. The 7 standalone module files are removed. Rendering now works.
Also fixes scene/planets.rs: upload_planets was allocating a fresh
ShaderBuffer (new handle) every frame, which re-triggered the
AsBindGroup RetryNextUpdate that the previous commit's startup pre-fill
was meant to eliminate. Now mutates the existing buffer asset in place
via set_data, keeping the handle stable.
The committed Phase 1 code rendered a black screen due to three
independent bugs, fixed here:
- black_hole.wgsl: dt was sized as |eye|/steps, so the full integration
only traveled |eye| units total — rays never reached capture or escape
and every pixel fell through to accum=black. Now dt covers
eye_dist + escape_r across the configured step count, and the escape
test uses the dynamic escape_r instead of a hardcoded 1000.0.
- render/plugin.rs: the planets binding was Handle::default(), causing
AsBindGroup to return RetryNextUpdate every frame and silently skip
the fullscreen quad's draw. Pre-fill a MAX_PLANETS-sized zeroed
ShaderBuffer at startup so the binding resolves immediately.
- render/plugin.rs: scale the fullscreen quad by half the window size
to match Camera2d's default WindowSize projection (1 unit = 1 px);
the old aspect-based scaling letterboxed the image.
- render/plugin.rs: add nudge_camera to work around Bevy 0.19 #24448,
where a static camera stops rendering after the first frame.
- planets.wgsl: rotate world-space planet centers into disk-local space
before the ray-sphere test (ray and centers were in different spaces).
README documents the renderer, run instructions (desktop + web), controls,
how the geodesic pipeline works, performance notes, and Phase 2 (Kerr) as
future work. Allow dead_code on BlackHoleParams since render_scale + spin
are intentionally reserved for Phase 2.
The quad mesh was built once at startup with the initial aspect ratio, so
widening the window afterward left empty edges. Add a FullscreenQuad marker
component and a fit_quad_to_window system that rescales the quad's Transform
to the live aspect each frame (only mutating on change).
Also remove the misleading 'Render scale' slider — it isn't wired to a real
sub-resolution render target in Phase 1 (the full-screen quad always renders
at window resolution). Lowering Steps is the real perf lever. Documented in
a code comment for future work.
The whole-file #import "file.wgsl" form (without ::) does not reliably
bring functions into scope in naga_oil/Bevy 0.19 — the renderer was
failing at runtime with 'no definition in scope for identifier:
ray_direction' (and rot_x, deriv, etc.), producing a blank screen despite
'cargo build' passing (shaders aren't compile-checked at build time).
Fix: add #define_import_path to each module file and import symbols
explicitly via namespace::name (the canonical Bevy pattern from the
shader_material_2d example). Verified the full shader pipeline compiles
and runs at runtime with zero WGSL errors.
This was a pre-existing bug masked by incomplete runtime verification in
earlier tasks; it is NOT specific to Task 15's grid work. All earlier
visual features (shadow, Doppler disk, Einstein halo, stars, planets,
grid) now actually render.
Add planets as ray-traced spheres inside the RK4 geodesic loop. A new
planets.wgsl shader tests each integrator segment against a storage
buffer of SphereData (center/radius/color/emissive) and composites hits
front-to-back alongside the accretion disk. Lambert shading with a fixed
light direction; emissive flag bypasses shading.
- assets/shaders/planets.wgsl: segment-sphere intersection + shading
- src/scene/: Planet component, upload_planets system, default planet
- black_hole.wgsl: wire planet_hit into the compositing loop
- geodesic_schwarzschild.wgsl: replace tuple return with Deriv struct
