1,411 words

Status: accepted calibration record · 4 Sep 2026. Camera-mode policy is superseded.

ADR-0037 governs the implemented Enhanced, Automatic and Manual modes. This record preserves the measured Natural, Composite and Direct baseline, including ACES diagnostics, optical calibration and resource ownership. Those response names below identify that baseline; they are not additional player modes. Explicit cinematic calibratedLight staging retains the measured ACES look. Ordinary photographic views share the neutral response and do not use Direct channel clipping.

Context

This record builds on ADR-0029: the scene already belongs to one chain, but exposure, focus and shutter have no effect on its output. A physically exposed sky and a luminance-only tone curve also change the production image: stars disappear beside a lit hull, and saturated highlights retain more color. The visual constraint is to preserve that familiar default while making the photographic responses deliberate camera choices.

Decision

One sensor chain reads the lens; Natural retains the production lighting calibration, ACES fit and integrated sky, while the other Composite presets and Direct collect physical stellar flux.

Surfaces use a solar-calibrated luminance scale of 30,000 cd/m² for an albedo-one surface. Stellar luminance and incident irradiance derive from luminosity, radius and distance. These are visible-light estimates from bolometric catalog values, not spectral photometry. Scene materials apply pre-exposure before the half-float target; offscreen reflectance bakes do not.
The WebGPU meter samples every fourth pixel in each axis into 64 log bins. A bounded asynchronous readback feeds the CPU percentile reducer. It excludes empty sky and meters the 40th–95th percentile of the remaining samples. Adaptation uses simulated time, with 0.4 s toward bright and 3.5 s toward dark. Rate, bright/dark range and display peak are preferences. A tightened range binds even with adaptation held. WebGL uses a fixed Composite calibration.
Direct uses the lens's EV. The photographic Composite presets add automatic gain without writing ISO. Natural holds the surface calibration, bounded by the comfort range, and retains the production dark-body lift and star stop-down. A script pins an EV offset from the surface calibration and explicitly opts into calibrated lighting where its reference staging needs it. The readout distinguishes those fixed frames from metered ones.
Natural uses the production ACES matrices and fit. Its sky retains the seventeen-magnitude visibility ramp, compensated for the sensor's total exposure so a scripted exposure offset does not erase the integrated stars. This is declared Composite processing, not a physical claim about the detector. Neutral uses the fit on luminance with uniform RGB scaling; Gentle and Crisp use that response with different HDR shoulders. Direct clips channels.
The pass carries two half-float attachments: radiance and velocity.xy with reciprocal view-space depth. Reciprocal depth represents nearby meters without overflowing on astronomical distances. Lens-derived defocus gathers near/far layers at half resolution and bypasses all draws for a sharp frame. Motion uses tile-max, neighbor-max and a depth-weighted gather, bounded at twenty pixels. The planetarium and documentation disable motion through their presentation stances.
Glare redistributes 1.5% of the image through six downsample/upsample octaves, without a threshold. Natural also retains the calibrated analytic Sun glow and horizontal streak: the faint solar point beside Titan cannot fund that broad production halo from its sampled image alone. The other responses use the sensor PSF alone for the core. Analytic ghosts and the scripted corona remain. The detector adds tick-indexed shot/read noise, lens-derived lateral color and Direct vignetting. White balance defaults to the unchanged D65 calibration.
The final output is opaque, with triangular dither on SDR. An extended WebGPU canvas uses P3 only when its configuration and the encoder can agree. Missing configuration reporting or a failed P3 configure retains sRGB. CanvasGamut owns the negotiated color space, reapplies the agreement after every canvas resize, and commits that live value after R3F configures output. Renderer diagnostics read the same value. Disposal removes the resize listener. Display headroom remains an authored maximum of two, reduced by the peak cap.
Every optical pass and the output quad register through the warm-up. Internal quads read plain textures; one dependency graph schedules the scene and each optical pass once per render call. The chain restores renderer state on a throw. engine.exposure and engine.sensorDiagnostics expose the live result. Each optical pass privately pairs its materials with their render targets; warming, drawing, resizing and disposal use those same records. The sensor calls warm(renderer) and reads outputTexture, without inspecting the pass's internal stages. Each pass retires its resources once.

Alternatives considered

Make the hue-preserving curve and physical star field the default. The curve passes its color measurement, but a pinned surface exposure removes the familiar star field. Automatic gain also takes a framed Bennu from the calibrated EV 14.61 to EV 8.11, making a dark asteroid nearly white. Natural retains the calibrated lighting and exposure as well as the curve; metered exposure remains an explicit photographic choice.

Subtract the direct core to measure a small glare skirt after half-float resolve. Rejected as a kernel measurement: the 400-unit point resolves to 399.787, a 0.053% frame error, but subtracting its 394-unit core attributes a 3.55% error to the six-unit skirt. The test measures the isolated kernel within 2% and the combined frame within 0.1% separately.

Switch to the half-resolution layer as soon as a blur exceeds half a pixel. Rejected: a 0.5625-pixel circle replaced an alternating 0/1 signal with 0.5, softening the reference hull. Coverage now fades through the subpixel interval. A large circle elsewhere in frame also cannot darken a nearer layer by reducing its gather coverage: normalized color and edge opacity have separate jobs.

Small circles use twelve equal-area iris samples through a four-pixel diameter; larger circles retain forty-eight. Both gather variants warm before use, share the same four targets and submit the same four passes. The half-pixel bypass, subpixel coverage fade and color normalization remain in force. At four pixels, the striped near/far fixture differs from the forty-eight-sample reference by at most 0.01172 in a linear channel, with a 0.00893% energy difference. The existing large-blur diameter calibration is unchanged. Six alternating sixty-frame drained batches measure the isolated defocus chain at 0.931 versus 0.376 ms at 1920×1080, and 2.259 versus 0.921 ms at 2880×1800 on Apple M5. These are optical pass measurements, not complete-scene frame rates.

Consequences

The default is visually close to production, not pixel-identical. Glare, noise and P3 still change the image. At headroom one, the Natural tone curve itself agrees with stock ACES to 1e-6 on the GPU. Neutral's measured hue drift for 3,000 K and 10,000 K sources across radiances 0.1–100 is below 0.00002°. The two-meter defocus calibration measures 36.318 px against the thin-lens 35.834 px, within the half-pixel gate. Paused pinned frames repeat exactly in the GPU test; automatic adaptation still receives asynchronous meter readings.

The three Enterprise portraits exercise held composition and close focus; each camera clears the shipped hull by the existing fifteen-meter test margin. Their thumbnails are vendored captures. Clearing chrome also hides the debug cinematic transport, so those captures contain only the picture.

Star sprites supply their instance position to both sides of the velocity calculation. The unit quad only describes the pixel footprint; using it as the previous position invents motion between a star and the render origin. The homepage's automatic phase sweep is sampled inside the observatory at render time, so pausing the simulation holds the camera too. Its keyboard context leaves game controls to the browser while keeping Settings and the keys sheet.

An offscreen WebGL 2 draw of a gray calibration plane through the full chain at four samples reads opaque [163, 165, 164, 255] with no GL error; that is a one-off check during verification, not a test in the tree. P3 is confirmed in Chrome 152. three configures the canvas lazily and forgets the configuration on every resize, so the canvas and encoder renegotiate from the canvas target's resize event. A failed P3 reconfiguration switches both to the restored sRGB declaration; a later successful resize can recover P3. Without that agreement, the first setSize after the renderer is built can return the canvas to sRGB while the encoder keeps writing P3 primaries.

The FFT diffraction tier and spectral emission attachment are not implemented. Neither are narrowband filters, a photo export/tether workflow, a spectral calibration, or automatic display-headroom discovery. Iris samples currently use the flight glass's blade pattern. The meter reduces on the CPU rather than in a second GPU kernel. These are open parts of the plan, not capabilities the response control claims to provide.