worldlightmapbake

Offline directional lightmap baker

bake/worldlightmap (Go package worldlightmapbake) is a public, CPU-only producer for the SDK's completed RNM prebake format. It accepts a validated compiled world, its existing receiver layout and original albedo images. It does not import the private engine or require OpenGL, WebGL or WASM execution.

For chunk receiver algorithm 2, pass the complete logical world and the selected layout to a world-bound Baker. Only selected receivers allocate atlas space; all ordinary physical neighbors remain shadow and diffuse transport geometry. Other physical components are excluded, including those reached through transformed portals. The producer preserves a selected space with no local light as covered black. Tests exercise a light and occluding solid outside the receiver chunk, and a transformed-space lighting boundary.

baker := worldlightmapbake.NewBaker(document, materials)
options := worldlightmapbake.Options{Samples: 64, Bounces: 1, Seed: 1}
for _, layout := range chunkLayouts {
    result, err := baker.Bake(ctx, layout, options)
    // Handle err, then encode/package result.Image with result.RGBMRange.
}

Use one Baker for repeated chunk or clock/weather bakes of an immutable world. Use baker.BakeEnvironment(ctx, layout, settings, options) for sun/sky layouts without a point-light recipe. Create a new Baker after editing world geometry, materials or source images. It retains only the last prepared transport scope, not every atlas. Package-level Bake and BakeEnvironment remain convenient one-call entry points to the same core implementation.

Samples default to 16, Workers to GOMAXPROCS capped at 64, and Bounces is explicit (zero means direct-only). Keep sample/bounce/denoise choices explicit in project policy; the core has no separate preview or experimental mode.

Material pairs the compiled numeric material ID with an image.Image of its original sRGB albedo. Missing nonzero opaque-surface materials fail explicitly. The caller must keep images and document/layout slices unchanged during the call. Use Result.Image as straight-alpha RGBM, without premultiplying its alpha, and use Result.RGBMRange when making an algorithm-2 offline prebake manifest. ReflectanceDigest independently commits the complete world's used decoded texture bytes and semantic material/UV assignments, including affine projections and triplanar weights. Unused materials, actors, ambient and material AO do not invalidate that digest. The separate prebake manifest also commits geometry, layout, selected lights and bake controls. ReflectanceDigestForLayout(document, layout, materials, options) returns exactly the producer's current scoped reflectance provenance, including its ordinary physical component and any bounded transport halo. Use this value to create a fresh manifest when reusing verified cached pixels.

The baker rasterizes actual semantic polygons into chart texel centres and uses a surface-area triangle BVH for double-sided ray visibility. It includes solid caps/sides and the opaque bands of partially open portals; door openings do not become blockers. Chart rectangles remain isolated during coverage dilation, and covered black samples retain nonzero alpha. Runtime recipe validation restricts the bake to one ordinary reciprocal physical component. Transformed portal transport is not supported.

The original Bake entry point uses the same point-light-only producer in environment worlds. Sky ceilings, horizon boundaries and outdoor upper seams contain no polygons, so they neither cast shadows nor require albedo images. Escaped transport rays receive zero incoming radiance: this producer does not bake directional sun, sky illumination or sky-driven bounce light. Those contributions remain live. Changing presets has no effect on bake pixels, although the canonical offline surface snapshot conservatively retains the environment for artifact identity. Sun selection in a point-light recipe is rejected; transformed-portal and ordinary physical-component restrictions remain unchanged.

Environment producer decision, version 1

Candidate SDK 0.0.12 adds the separate BakeEnvironment(ctx, document, layout, materials, settings, options) producer cpu-environment-rnm3@1. It leaves the point-light producer and its pixel fixtures unchanged. The input layout has no point-light recipe; every authored time/weather variant shares its exact chart coordinates. EnvironmentInputDigest binds those coordinates, complete static transport, used original albedo, time/weather and deterministic sampling.

This producer bakes static directional sun and sky diffuse through actual openings. Opaque geometry casts hard sun shadows. Hemisphere rays that escape the ordinary physical component receive the resolved directional sky fill; blocked rays receive no invented ambient term. Optional diffuse bounces use the existing albedo, RNM, deterministic path sampling and indirect denoiser. The cloud amount scales sun diffuse by 1-cloud_amount, so clear and overcast endpoint maps can supply continuous weather by blending decoded linear RNM coefficients. RGBM bytes themselves must never be interpolated directly. The CLI defaults to clear-only time images (weather_variants=false). It calls this producer with cloud amount 0 once per authored clock; cloud cover then changes a renderer brightness multiplier while retaining the baked shadows. Paired weather baking is an explicit authoring opt-in.

Closed rooms stay covered black unless light enters an opening and reflects. Translucent liquid materials (opacity < 1), including static window glass, neither block opaque visibility nor receive opaque RNM pixels. This initial producer treats their transmitted lighting as unobstructed; it does not model glass tint, refraction or partial absorption. Point lights, authored ambient, actors and specular remain runtime contributions. A matching renderer must suppress live sun/sky diffuse on covered static surfaces to avoid adding it twice, while retaining the runtime terms.

The existing page, sample, bounce, worker and material bounds remain in force. Environment rays must be unbounded (MaxRayDistance = 0): an artificial cutoff cannot prove that a ray escaped to the sky. Receiver selections retain complete ordinary physical-component geometry as transport context; transformed spaces cannot share one environment bake. A 1024 RNM map still occupies 12 MiB decoded, so consumer caches must budget maps explicitly rather than retain every clock and weather combination.

Primary direct lighting retains the current squared Half-Lambert RNM contract and squared falloff in normalized squared distance. Bounced transport instead uses Lambertian cosine response, cosine-weighted hemisphere sampling and linear albedo reflectance clamped per channel to 0.95. UV lookup uses nearest original texture pixels with the compiled projections and weights. Legacy unmapped caps use world XY and sides use edge distance/Z with one repeat per world unit. Material normal, height and AO do not alter transport in this slice.

Each sample traces one bounded path through up to four diffuse reflections. Selected direct lights are ray tested at each hit; reflected contributions are accumulated into the receiver's three directional RNM coefficients. The coefficient distribution preserves geometric neutral irradiance. No dense surface-to-surface matrix, artistic constant fill light or ambient term is used. Producer revision 2 reserves a conservative geometric-series energy bound over spatially overlapping light influences. A weighted interval sweep computes the maximum possible overlapping source energy independently along X, Y and Z; the smallest of those three upper bounds remains conservative for every sphere. Outward interval and accumulation rounding preserve that bound. A thousand distant lamps therefore retain approximately the range of one lamp, while a thousand overlapping lamps reserve their combined energy. The atlas still has three coefficient tiles at the receiver's fixed resolution. New producer names end in @2 and explicitly bind their RGBM range into the offline identity; @1 manifests remain verifiable using their original global-sum range and exact identity bytes. Source selection, shadows, path sampling and accumulation order retain their existing semantics.

Options.MaxRayDistance defaults to zero, selecting unbounded revision-4 transport. A finite positive value at most worldlightmap.MaxBakeRayDistance (1,000,000 world units) selects a distinct revision-5 producer. A source beyond that distance contributes zero to both primary and bounce-hit lighting; sources inside the limit keep their original radius falloff. Every diffuse bounce segment is also limited to that distance. This intentionally truncates transport and can reduce distant bounce light. Denoising visibility segments stay within their receiver charts and retain their barrier tests independently of the limit.

BoundedInfluenceScope is the authoritative conservative dependency selection. For b bounces and distance D, derived surface polygons overlapping receiver bounds expanded by (b+1)*D remain transport geometry. Sources must reach receiver bounds expanded by b*D, using influence radius min(source.Radius,D). Numerical guards cover ray-origin and raster rounding. The local source set controls RGBM range, while the complete recipe continues to classify completed static sources; scoping never returns an excluded steady source to runtime light slots. Bounding boxes can overestimate dependencies, especially for large receiver charts.

BakeInputDigest(document, layout, materials, options) hashes the present pixel dependencies: producer, sampling/filter controls, distance, local RGBM range, source records, chart packing, actual derived polygons, UV/material assignments and used decoded albedo pixels. It omits global geometry digests, unrelated bindings and global sector/solid index offsets. Used local geometry and source order are retained. Worker count does not affect it. Thus a cache from hundreds of commits ago can be safely looked up by recomputing current content; a Git diff or last-build snapshot is unnecessary. Remote edits outside the halo preserve the key when local receiver packing remains the same. The cache must verify its stored QOI integrity and regenerate current layout/manifest identities before packaging. The digest identifies reusable pixels, not a mountable old artifact. Zero distance conservatively retains the full physical-component dependencies.

Recipes accept up to 4,096 authored static sources independently of the 50 live direct-light limit. Named-source lookup uses a bounded index. Above 32 selected sources, a balanced influence BVH skips lights whose conservative radius boxes do not contain the receiver or bounce hit. Its bounded bitset iterator yields candidates in recipe order, preserving deterministic floating-point sums and existing pixel fixtures without per-hit allocation. Smaller recipes keep the linear walk. This accelerates spatially separated sources; physically overlapping lights still require their actual build-time visibility work.

The sparse source benchmark runs one direct RNM query and one bounce-hit Lambert query against the same source set, with no occluders. On Linux amd64, AMD Ryzen 7 5700U, pinned Go 1.27.1, three 300 ms repetitions produced these median times; every variant allocated zero bytes and zero objects per query pair:

Command: GOCACHE=/tmp/karty-engine-go-cache mise exec -- go test ./bake/worldlightmap -run '^$' -bench '^BenchmarkSparseSourceLighting$' -benchmem -benchtime=300ms -count=3 -timeout=9s. These are CPU source-query measurements, not whole-bake timings or GPU/frame-rate predictions. The 24-source Hangar recipe keeps the small-list path.

Samples default to 16 and accept 1–256. Bounces are explicitly 0–4; zero requests direct-only output. Workers default to GOMAXPROCS capped at 64, with an explicit range of 1–64. Seed zero is valid. Per-texel/sample seeds make image bytes independent of worker scheduling. Context cancellation returns an error and no partial image; reflectance hashing and raster/path stages check cancellation regularly. Reflectance hashing validates at most 16 million input pixels and 16,384 material records. Receiver positions occupy only chart receiver rectangles, at 32 bytes per reserved receiver texel. Dilation reuses scratch sized to the largest padded chart: one distance byte and one native-int queue slot per chart texel (9 bytes on 64-bit targets, 5 on 32-bit targets). Output remains 12 MiB for a 1024 page; BVH/triangle geometry, tile scheduling and the caller's decoded images are accounted separately. Unused page space does not allocate receiver or dilation scratch. No per-ray scratch allocation is needed. mise run bench covers 512 and 1024 pages with the same small receiver fixture, plus direct and bounced work.

This first producer uses hard point-light shadows and finite-sample diffuse transport. Low samples can produce visible noise; use more samples for final assets. It has no soft area-light sampling, spatial actor probes, emissive materials or moving shadows. Its primary direct lobes preserve Karty's current Source-inspired approximation rather than reproducing Valve's RNM equations exactly.

Tests verify coloured reflection into a direct shadow, absence of reflected energy from black surfaces, no leakage across a closed barrier, partial-portal blocking/opening, UV-dependent reflected colours, the neutral direct contract, energy bounds, worker determinism and cancellation without partial output. Six golden pixel/reflectance hashes pin the original one-source transport fixtures across producer revisions and internal refactors. A separate manifest fixture preserves the exact revision-1 identity and pins revision 2 separately.

Default material decision, version 1

This additive behavior targets the next Go module release after v0.0.8. Compiled sector surfaces and material atlases already accept ID 0; the baker now accepts it as well. If omitted from materials, it resolves to a bounded one-pixel opaque sRGB (192,192,192) albedo, matching the CLI's default atlas. That fixed four-byte fallback is separate from the caller's input-pixel budget. An explicit Material{ID: 0, Albedo: image} supplies the caller's default albedo and follows the same input bounds and duplicate checks as every other material. Missing nonzero materials still fail, including direct-only bakes.

ReflectanceDigest hashes the resolved default pixels only when a receiver uses ID 0. Thus omitted and explicitly supplied identical default pixels have the same identity; changing a used explicit default image invalidates the bake. Previously valid nonzero-material worlds retain identical pixels and digests. The candidate uses canonical typed binary for reflectance assignments and bake identities; runtime JSON has no fallback. Artifacts and bake caches must be rebuilt. Producer identifiers, sampling, pixels and capability names retain their existing semantics; the versioned candidate format decision owns the serialization change.

Indirect denoising

Options.Denoise accepts off (also the empty SDK default), low, or medium. Low applies two spatial à-trous passes (strides 1 and 2); medium applies three (1, 2 and 4). This is an offline pure Go filter, with no new module or native dependencies and no runtime frame cost. Direct light is kept separate and added after filtering. Albedo, normals and material textures are never filtered by this pass.

The filter uses linear floating-point RNM coefficients before RGBM encoding and gutter dilation. All three RNM lobes share weights based on indirect colour and estimated sample-mean variance. It stays inside covered chart texels; adjacent geometry visibility segments prevent smoothing through solid bases and thin walls within a merged chart. Wider passes walk those visibility edges and cannot skip coverage holes. Worker count does not change pixels. Cancellation never publishes a partial atlas.

Additional retained storage is 76 bytes per reserved receiver texel, plus 37 bytes of scratch per concurrently processed chart texel, bounded by the existing MaxTexels limit. Stats expose DenoiseDuration and DenoiseRays; the latter includes the short visibility segments in total Rays. Presets are versioned producers; off retains the original exact producer pixels.

This is a spatial estimator, not reconstruction of missing transport. It may soften small indirect-light features, leaves some residual grain, and does not reconcile lighting across separate charts. Use more samples for small bright sources or difficult contacts, and compare against denoise off.

Portable SIMD acceleration

The next Go module release after v0.0.8 uses Go 1.27's portable simd package for float64 double-sided triangle intersections. Build consumers of bake/worldlightmap with GOEXPERIMENT=simd; the SDK, CLI and engine root mise environments set this automatically. Remove that gate when the pinned Go toolchain enables SIMD by default. Format-only imports need no experiment. There is no CGo, assembly maintained by Karty, or new module dependency.

The BVH keeps precomputed triangle edges in contiguous numeric columns. Each packet tests one ray against as many triangles as the platform's vector width permits; partial leaves exclude neighbouring triangles and padding. Shadow and denoiser visibility queries stop at the first blocker; diffuse paths still find the nearest surface. Four-child box packets retain float64 division to preserve grazing-ray classification. Float64 tolerances, unfused triangle arithmetic and sample sequences stay consistent across native and emulated execution. Exact hit ties use the lowest semantic surface index, independently of tree order.

Hardware selection and unsupported-platform emulation belong to Go's SIMD package. There is one production triangle kernel; the old scalar arithmetic exists only in tests for comparison. mise run test-simd-emulated forces Go's fallback with GODEBUG=simd=0. Existing producer pixel hashes also run in the actual WASM and native 386 tasks. A leaf has at most eight triangles and ray queries allocate no heap scratch. Packed triangles retain nine float64 columns and two int32 primitive fields, plus less than one vector of column padding. A cold median tree and one int32 per-triangle lookup preserve existing grazing/overlapping-surface behavior. Construction triangles are released after packing.

Diffuse paths stop immediately when all three throughput channels become exactly zero. An absorbing hit cannot contribute later lighting, so this preserves producer pixels and cache identities while avoiding further bounce and shadow queries. Small nonzero throughput still follows the existing path; no stochastic termination is enabled. Ray counts can fall without an increase in Mrays/sec, so compare completed bake time as well as query throughput.

For a same-process scalar-oracle/SIMD comparison:

mise exec -- go test ./bake/worldlightmap -run '^$' -bench BenchmarkRayTrace -benchtime=1s -count=3 -benchmem

Measure full karty render bake runs separately: traversal, texture reflectance, rasterization, denoising and output work can limit the total speedup.

Measurement, 2026-10-06

Linux amd64, AMD Ryzen 7 5700U, pinned Go 1.27.1, default 128-bit hardware SIMD. Three repetitions of BenchmarkRayTrace with -benchtime=500ms had median times of 770.1 ns for the scalar nearest-hit oracle, 675.2 ns for SIMD nearest-hit, and 660.4 ns for SIMD visibility. All report zero allocations.

An old/new/new/old CLI comparison used the same Roman court geometry (now retained in the CLI testdata/roman-court fixture), 8 workers, 430,772 receiver texels, and medium denoising. The equivalent invocation with the current CLI is:

karty render bake --workers 8 --samples 32 --bounces 2 --denoise medium

This is a 23.7% reduction in complete CPU bake time, including denoising and output. Both implementations issued 57,263,456 rays and produced identical complete prebake manifests and encoded image bytes (image SHA-256 23600e5eeeb3b99d3841f13c348de1cdc12563825ef5c9b225d2877a85db6865). The four-sample, one-bounce comparison also produced identical bytes; warm runs remained roughly two seconds, and its initial scalar timing was an outlier, so no precise draft-bake speedup is claimed.

These numbers measure this CPU and scene, not Mac/ARM performance, WASM bake speed, browser graphics, or frame time. Native 128/256-bit SIMD and forced emulation checks cover ray equivalence, leaf tails and original pixel hashes. Go's default hardware selection is unchanged; the optional 256-bit check used GODEBUG=simd=+256 only on this AVX2-capable test machine.

Per-bake precomputation

The baker prepares the exact hemisphere frames for surface and receiver-chart normals once. A bounded table of at most 256 initial samples stores their radial and normal magnitudes, azimuth increments and seed factors. Per-receiver origin and seed setup runs outside the sample loop. Point lights retain squared radii, and fixed squares use multiplication rather than the general power function. The random sequence, RNM basis, light response and sampling controls stay the same.

Sampled albedos are cached by material ID for one bake. Existing image.NRGBA images, including strided subimages, are borrowed unchanged. Other representations are converted once using the existing exact decoded-pixel conversion. Texture bounds and diffuse-reflectance lookup values are prepared ahead of sampling. There is no texture filtering, rescaling, lighting interpolation or cross-bake cache. Inputs remain caller owned and immutable until the bake returns; original source pixels still determine the reflectance digest.

Only bakes with diffuse bounces allocate frames, sample tables or prepared albedos. Frame storage is 48 bytes per surface and receiver chart, initial samples occupy at most 8 KiB, and converted albedos add at most 4 bytes per input pixel (64 MiB under the existing 16-million-pixel input bound). Converted textures are shared by surfaces using the same material. Cancellation never publishes a partial bake.

Measurement, 2026-10-06

Same Linux amd64 Ryzen 7 5700U, Go 1.27.1, 8 workers, medium denoising. Both binaries were warmed, then measured in before/after/after/before order. The baseline already includes the portable SIMD and early-exit visibility changes.

The setup/material changes alone reduced the full 32-sample bake by 3.5% (16.345 s to 15.771 s). The small one-worker fixture improved about 12%, so it was not a reliable predictor of the full scene's gain. A full-scene CPU profile then identified ray-box interval tests at 57.6% of sampled CPU time, including substantial general math.Min/math.Max helper overhead. The interval loop now uses comparisons for validated finite coordinates; exact division, parallel-axis tolerances, traversal order and hit classification are preserved. Signed-zero interval endpoints cannot change its Boolean result.

Fresh paired measurements of the combined setup and ray-box improvements:

For the 32-sample case, average end-to-end throughput rises from 3.466 to 4.093 million rays/second, about 18.1%. Both perform 57,263,456 rays and produce the same complete manifests and encoded atlas bytes. Preparation, denoising and output costs are all included. These are native CPU bake measurements, not WASM performance or frame-time results. The initial sampling and light-response formulas, retro texture filtering, quality controls and producer IDs are unchanged.

A separate run at 36 samples and two bounces completed 64,253,080 rays in 16.040 s, including medium denoising. That is 12.5% more samples within the previous 32-sample average budget of 16.520 s. It is one confirmation run, not a paired quality benchmark; the authored sample defaults remain 4 samples / 1 bounce.

The SDK 0.0.13 daylight-bright preset supplies four times daylight sky radiance, with the same sun and sky appearance. Existing presets retain their pixels. Selecting it changes environment input/surface identity and the RGBM range, so old condition maps cannot be reused. It requires the cartridge capability world/environment-skylight@1.

Shared transport preparation decision, version 1

Public Go module v0.0.15 adds NewBaker(document, materials), returning a world-bound Baker with Bake(ctx, layout, options) and BakeEnvironment(ctx, layout, settings, options) methods. The original package functions remain available. Every call performs complete world/layout/material validation and computes current provenance before using shared preparation. Keep the document, material records and source images unchanged until the Baker is discarded. Create a new Baker after editing any of those inputs.

A Baker retains only its most recent transport scope: prepared triangle data, BVHs, original-albedo conversion, surface reflectance mappings and normal frames. Included surface IDs plus sampling/reflection requirements distinguish scopes within its fixed immutable world. Bounded halos, physical components and point/environment opacity selection cannot silently share incompatible geometry. The CLI creates one Baker per chunked level build, including clock/weather variants, and releases it afterwards. Concurrent SDK calls are safe; receiver rasterization, light recipes, samples, filter buffers and output remain per-call. This is a build-only additive Go API, with no guest API, wire, manifest, producer or pixel-cache identity change. It does not retain every chunk or clock image.

The ray BVH uses stable 16-bin surface-area splits, with median fallback for unsplittable or deep partitions. Small leaves spanning multiple surfaces split down toward two triangles; a single surface's fan stays together at up to eight triangles. A balanced median partition handles degenerate/deep splits within that same construction tree. The temporary binary construction nodes occupy 64 bytes each and are released after packing the only retained BVH: BVH4. Six contiguous four-value columns hold each internal node's float64 child bounds, with compact int32 references and separate leaf ranges. SIMD tests one ray against the child boxes using division rather than approximate reciprocals. Conservative bounds include the triangle tolerance. Nearest hits choose the smallest distance and then the lowest semantic surface index at exact ties; any-hit visibility stops at its first blocker. Ray upper endpoints remain strict. No old tree, replay traversal, legacy rank or legacy-order mapping is retained. Receiver workers claim 16-by-16 chart-confined tiles and accumulate ray counts locally. Sampling order within each texel, RNM coefficients, denoising and encoding stay unchanged.

Point-light evaluation computes the receiver epsilon once per receiver or bounce hit and reuses it for source-distance checks and shadow-ray offsets. The traversal guard still comes from the shifted ray origin. A point with no point lights returns after any sun contribution, so sun/sky bakes avoid the unused point-light setup. The single geometry BVH is shared by point-light, environment, bounce and denoiser ray queries. The separate light influence index selects candidate sources; it does not duplicate triangle geometry.

The consolidated point-light producers are revision 4 (unbounded) and revision 5 (bounded), including their low/medium denoising names. Environment baking emits cpu-environment-rnm3@2. Producer/cache identities change because removing old-tree pruning can change edge or coplanar winners. Old completed artifacts remain readable through format validation; old tracing implementations are not retained. See the single-BVH producer decision.

Experiment results, 2026-10-14

The temporary Go-overlay runners, sampler/triangle/shadow templates and research mise tasks were removed after consolidation. These notes retain their findings; none of the rejected proposals is a selectable bake option. Measurements used Linux amd64, Ryzen 7 5700U and Go 1.27 float64 SIMD. Timings are fixture-specific, not universal throughput or browser/device results.

Adopted work

Shared immutable transport preparation, contiguous triangle columns, float64 SIMD intersections and BVH4 boxes, 16-bin surface-area splits with smaller mixed-surface leaves, 16-by-16 chart jobs and worker-local counters form the production path. Receiver epsilon reuse avoids repeated point-light setup. Absorbing paths stop when every throughput channel is exactly zero; this preserves pixels and removed roughly 78–86% of queries in absorbing-room fixtures, with no measured courtyard benefit.

BVH4 improved complete courtyard bake time by about 24% over the preceding binary SAH traversal. The later leaf tuning reached 8.12–8.18 Mrays/s at eight samples/two bounces and 9.88–9.92 at 32 samples/four bounces. Higher-sample 63k/127k-triangle worlds improved about 8.5%/4.7%; fresh low-sample dense bakes were 3.5–4.6% slower and needed more preparation allocation.

Receiver epsilon reuse reduced courtyard bake time by 7.0% at eight samples/two bounces and 9.7% at 32 samples/four bounces. Three complete high-sample SDK Bake calls reached 10.72–10.92 Mrays/s with identical atlas bytes and 114,513,581 queries. Their mean duration was 10.58025 s, versus 11.71734 s before epsilon reuse. These calls include preparation, rasterization, tracing, medium denoise, RGBM encoding and dilation; they exclude compilation, QOI and file I/O. The large two-light fixtures did not establish a general shadow-setup speedup.

Traversal and layout proposals that were not adopted

Replacement-only triangle packets showed an incoherent-ray synthetic cache benefit at 65,536 triangles, but only about 1% improvement in validated 63k/127k-triangle cylinder-world bakes. The courtyard was neutral. Global triangle columns therefore remain. Forced 256-bit SIMD was slower on this CPU; Go's default vector selection remains in use.

Near-first child sorting helped larger geometry but slowed the courtyard; cheap directional ordering had no consistent complete-bake advantage. Reciprocal slab arithmetic and precomputed sign flags did not improve the measured bake. Box tests retain division and conservative float64 bounds.

Eager/lazy shifted-origin and guard caching provided no clear additional courtyard gain over epsilon reuse. Dedicated opaque visibility traversal was about 2.65% slower in the initial courtyard screen. The generic nearest/any-hit kernel was retained during those measurements. The later consolidation removes its original-median replay and duplicate tree, using a new producer identity. Independent exhaustive scalar, strict-endpoint, packet-tail and reordered-child tests remain; the old BVH oracle was removed.

Sampling and work allocation proposals that were not adopted

Unencoded linear RNM was compared with denoising off, multiple seeds and independent high-sample references. Prepared-query timings exclude bake setup, rasterization and encoding. Full scrambled Sobol increased one-bounce red-ceiling MSE by 1.66–2.02 times. Keeping the existing first hemisphere and using Sobol only at later bounces preserved one-bounce pixels and reduced four-bounce white-room MSE by 11–23% at 32–256 samples, with about 5% transport overhead. Small-source results were mixed, so the existing sampler remains.

A 50/50 mixture of cosine sampling and uniform area sampling of an actual clipped reciprocal opening reduced zero-bounce doorway MSE from 0.008414 to 0.000907 at eight samples (9.28 times), and 4.95 times at sixteen samples. Independent guided/cosine references agreed to MSE 4.74e-8 and the blocked door stayed black. This small constant-sky fixture had four receivers. With four white-surface reflections, guidance improved error/time at 8–32 samples but regressed at 64: MSE was 1.82 times the current sampler while query time was 0.67 times. Cosine support and the complete area-to-solid-angle mixture density were retained. First-segment guidance needs up to twice the current RGBM bound; guiding every segment needs a larger bound and a separately versioned producer.

A discarded 16-sample pilot chose independent final 16/64-sample texel sets, with a neighbour-promotion variant. It lost on illuminated fixtures once pilot cost and error were counted; exactly dark fixtures used about half the fixed-64 work. Existing luminance variance was a rough classifier, not a certified RNM error bound.

Russian roulette after two reflections saved about 28–36% of queries on small illuminated four-bounce fixtures but increased MSE enough to lose error/time efficiency. Starting after three reflections was less disruptive and still inconsistent. Independent survival coins preserved direction streams and surviving future contributions were divided by their survival probability; already accumulated light was unchanged. Neither roulette nor adaptive budgets is enabled. Finite-sample weights, a new RGBM bound/producer/cache decision and any increase to the current four-bounce cap need separate qualification.

Build-only shadow geometry decision, version 1

Candidate SDK 0.0.13 adds Options.ShadowTriangles [][3]world.Vec3 to both point-light and environment CPU producers. These optional triangles are opaque, double-sided, black absorbing transport geometry. They block direct visibility and terminate diffuse paths, including sky paths, without becoming receivers, requiring albedo materials or changing reflectance provenance. This lets a compiler bake coarse vegetation grounding from deterministic species/seed proxies without adding live shadow geometry or per-plant runtime work. The original compiled world and receiver layout remain authoritative.

MaxShadowTriangles is 4096. The complete list is validated before bake scratch or acceleration preparation: every coordinate must be finite and within world.MaxCoordinate, and squared triangle cross-product magnitude must be at least 1e-28. Excess, degenerate and invalid geometry reject the whole call. Callers must keep this slice unchanged during each call and supply proxies in the selected receiver component's ordinary physical space; proxies do not transport through transformed portals or provide automatic sector ownership. Existing whole-world validation, receiver confinement, sampling, material, allocation and cancellation bounds remain in force. Proxy triangles add bounded BVH preparation/storage; they allocate no receiver atlas or reflectance image.

Both BakeInputDigest and EnvironmentInputDigest append the versioned opaque-shadow-triangles@1 count and ordered float64 coordinates only for nonempty geometry. Moving, adding, removing or reordering a proxy invalidates pixel cache identity. Nil and empty inputs append no bytes and retain existing producer strings, digest inputs and pixel fixtures exactly. A shared Baker hashes the geometry separately from its immutable source world and replaces its one cached transport preparation when geometry changes, including edits to a reused caller slice. Identical geometry still shares preparation across receiver and clock/weather calls. The returned Stats.Triangles includes proxies, while receiver count, RGBM bound and ReflectanceSHA256 retain their source-world meanings.

This is an additive public offline Go API decision. It adds no compiled-world field, guest command, manifest capability, RNM layout, prebake encoding or runtime shading requirement. Existing artifact readers remain compatible. Wind, cutout alpha and transmission are not simulated by these opaque proxies; consumers should use coarse, sparse geometry and report that approximation.