solvecraft/docs/render-cpu-general-glass.md

CPU general Glass with indirect lighting

The published implementation is in commit 00edaf6a, source tree 38a7f0af. Full CI7 ran the renderer, document, engine and UI tests; the renderer passed 744 tests and all 433 Render workspace tests passed. Full CI7 remains red in other listed tests. The native Sample Plate evidence used opaque geometry, so it does not qualify a native Glass image. See capabilities and qualification.

The Glass follow-up in this change prepares and retains the complete general-Glass certificate for Check and Start. The editor merges saved and prospective targets without saving the draft; checked file HDR capture keeps those stable identities. See PR #11087 for this follow-up's exact source and qualification; the historical CI7 and opaque native results above remain separate.

CPU Indirect lighting (options.path) can consume up to eight stable saved or exact captured optical assignments. It requires explicit progressive camera sampling, one primary attempt and traced shadows. Path depth remains 1–16 and counts shaded opaque surfaces, optical boundaries and active scattering events. An endpoint beyond this finite limit contributes no further radiance; the limit is not a convergence guarantee. There is no default, size, camera, profile or method change. Fixed saved choices require explicit adoption through Sampling or the CPU Glass transport section. GPU-only Fresnel strata and the GPU nested-media flag are not CPU controls.

Without path, the original planar sheet and strict axis-aligned box consumer, deterministic Fresnel branch tree, camera stream, raw pixels and cumulative work remain unchanged. Existing saved closed_mesh intent still lowers to that box subset in this route. With path, closed_mesh instead selects the new certificate and continuation. Explicit closed_box keeps its existing strict shape check. No saved optical number, engineering geometry, physical material or mass changes.

The new certificate checks the actual captured f64 triangle positions. It does not trust a closed flag, tolerance-weld seams, move vertices or repair triangles. Signed zero shares an exact coordinate identity. Closed targets require exactly two oppositely wound faces per edge, one connected shell and one connected circular link at every exact coordinate vertex. A pinched vertex fails the link check even when edge incidence and global connectivity pass. Adaptive exact predicates reject intersection, non-topological contact, coplanar overlap and self-intersection. Coordinates must be zero or have magnitude at least 2^-128 mm within the existing 1e12 mm world bound; this conservative predicate domain prevents underflow rather than snapping small coordinates. An exact signed dyadic volume proves outward orientation using a fixed 12-limb integer accumulator, not rounded volume or an approximate ray-crossing witness. Imported or computed seams that are not exactly closed refuse with their authored data intact.

Separated bounds use the existing fast disjoint test. Overlapping selected closed bounds need a conservative nesting proof: each enclosing parent is convex and all child vertices are strictly behind every outward parent plane. The immediate parent is independent of body order; touching or crossing optical boundaries refuse. Nonconvex nested parents remain unsupported. Thin sheets retain their existing planar certificate and may not overlap another selected optical bound. Optical contact with analytic rectangle emitters or point/spot source positions refuses explicitly. An ordinary opaque material wins an exact model-hit tie with Glass, preserving an opaque contact surface instead of stepping past it.

Each perspective primary ray classifies its physical eye or sampled lens pupil before any opaque hit or analytic emission. Exact half-open crossings initialize the ordered nested medium stack; Beer absorption starts at that physical origin, including opaque or emissive first hits inside Glass. Orthographic rays retain their established before-bounds origin. A conservative bounds fast path avoids classification work and allocation for unambiguous outside rays. Ambiguous samples use one reusable prepaid state and a cooperative BVH scan. An actual boundary at the origin or in the eye/pupil-to-near-clip segment refuses explicitly; endpoints outside a small object do not bypass the segment check. No camera, ray or boundary is moved to manufacture an outside origin. The single path carries an eight-entry medium stack. Every segment uses the previous physical surface point for Beer absorption, before any numeric launch adjustment. Geometric winding chooses entry/exit and verifies the certified parent. Smooth captured optical normals are used only if both reflected and refracted directions stay in the correct geometric hemispheres; otherwise the original geometric normal is used. Snell, exact unpolarized Fresnel, total internal reflection and the radiance IOR ratio are reused from the original f64 optics. On-boundary continuations exclude only the producing body's exact source plane; they do not consume a short segment with a finite normal offset.

Each Fresnel event chooses one counter-sampled delta branch. It creates no ray tree, extra image plane or extra camera sample. Its domain uses the existing per-pixel, per-ordinal, per-depth identity, including independent adaptive pilot seeds and the original production seed. Reflected or refracted paths give directly reached HDR or rectangle emission full delta weight. Opaque hits retain the complete existing Effects→Maps→Aging closure before every BRDF/PDF and denoising guide use. Maps remain unsupported on an optical target; unrelated opaque maps and aging remain active.

Thin-sheet transmitted shadows have an ordered closest-interface continuation. They apply RGB parallel-sheet Beer/Fresnel transmission and carry the transmission branch probability into the original light/BSDF sample-count MIS. If the remaining finite depth cannot realize the competing BSDF chain, light weight is one. Finite area connections retain geometry-wins-endpoint ties; point/spot connections retain their existing excluded-source endpoint. Closed refractive shadow connections are not implemented: authored shadow_transmission:true on a closed target refuses instead of silently using a straight-line approximation. Closed Glass still reaches HDR, area and mesh emission through actual sampled dielectric/GI paths. Point-light caustics and combining Glass with external participating volume remain unsupported. The independent bounded Atmosphere is an opaque-Path route; it does not erase those limits. No rough BTDF is added because the existing GPU Glass consumer is smooth delta transport.

Certification is cooperative: exact edge merge sorting, connectivity, vertex links, one face-pair contact check, determinant term or containment predicate run under the existing cumulative work and cancel token. For N selected closed triangles its prepaid peak is sizeof(Task)+N*(sizeof(Face)+6*sizeof(Edge)+3*sizeof(usize)+3); all vectors fallibly reserve exact capacity. The 12-limb volume state is in sizeof(Task). This change additionally bills one shared proof slot by sizeof(CertifiedOptics), including its Vec header; its outer Vec header remains in the existing fixed metadata allowance. The ready cursor's allocation moves into the image task, with its prior work charged once. Its source-specific qualification is recorded in PR #11087. Source-plane cursor state is billed by the actual existing pixel and path slot sizes. No cap or static guard is raised. Contact and containment can require quadratic work, so a valid large mesh can exhaust its selected profile without producing a partial certificate or image. The fixed 4096-byte host slot and all frame/metadata bounds remain compilation gates. The inside-origin increment adds one conditional exact-capacity state slot, prepaid by its actual size; its fixed BVH stack visits one node or primitive per charged step and is reused across camera samples. It adds no image plane, additional ray, random draw or new cap.

Renderer regressions cover exact contacts, pinched vertices, rotated and arbitrary closed meshes, outward sign under large-coordinate cancellation, nested physical Beer lengths, independent per-ordinal Fresnel paths, real HDR and area endpoints, opaque contact material, thin-shadow/BSDF MIS with 1/4/16/64 source samples, every PNG prefix, worker quanta, raw EXR identity, adaptive pilot discard, cancellation, work exhaustion, physical inside-eye and per-pupil nested Beer, first opaque and analytic emission, exact boundary/near-clip refusal, initial-medium TIR, saved/runtime capture and the real pointer preset→Render→Save→ reopen workflow. The published source's tests executed in CI7; its exact counts and failures are linked from the qualification record. The mathematical source reviews supplement those tests. A native Glass scene and image-quality acceptance remain pending; the opaque native fixture supplies neither.

Subsurface continuation

The selected-medium continuation and its mathematical, producer, saved/runtime and pointer-UI regressions are part of the published source tested by CI7. Native Glass/scattering visual acceptance remains pending. It consumes the existing saved OpticalMaterial.subsurface for a selected certified closed CPU Path medium. The scalar collision rate is per millimetre; HG anisotropy uses propagation directions, so positive values favor forward scattering. Beer absorption stays separate. There is no RGB scattering albedo, new rough BTDF, heterogeneous density or general external-volume consumer in this increment.

Each actual inside-eye/pupil or post-surface segment samples free flight from the physical segment start, before its closest model or analytic-emitter endpoint. The primary near clip does not shorten that optical length. A collision consumes one finite Path vertex, validates its actual f64 position against the current certified nested chain and launches one HG direction. It adds no ray tree, image plane, new sampler seed or cap. No-collision events retain their probability in endpoint MIS; contained direct-light connections use matching scattering survival and Beer absorption. No next-event connection through a refracting boundary is invented. Independently authored closed transmitted shadows stay refused.

Active scattering requires Path and disabled surface denoising, matching the existing GPU random-walk compatibility domain. Thin sheets and Path=None retain an explicit refusal, including authored inactive options on the legacy route. With Path, Some(rate0) stays authored and adds no random draw, continuation work or pixel/PNG/raw change. The pending collision/source-plane metadata is charged through actual sizeof(State), and captured optical storage uses actual element sizes; this is not a promise of unchanged metadata admission bytes across builds. The existing prepaid classifier slot is reused, with a debit before each scan or transition. With active scattering, opaque bounce and direct-light rays also start at the physical surface: a charged one-primitive-at-a-time range walk reuses the existing exact source-plane cursor. This avoids consuming collision/absorption distance through a normal offset; inactive options keep the old launch path. Nonrepresentable event positions or collapsed source/endpoint-plane positions refuse instead of being moved by an epsilon. All existing profile, 4096-byte host-task and 1024-byte frame-metadata limits remain unchanged gates.

Regressions cover independent exponential survival, HG CDF/density mass and first moment, per-ordinal first-hit eye/lens/nested Beer and slab walks, contained light/BSDF survival MIS at 1/4/16/64 source samples, inactive identity, white furnace, every complete prefix, quanta, raw/alpha/finishing/adaptive pilot separation, cancellation/work exhaustion and saved/runtime/pointer save/reopen. Their execution is covered by the published renderer suite, not by the opaque native Sample Plate capture. See PR #11087 for the Glass follow-up's exact source and executed qualification scopes.