solvecraft/docs/render-materials/BATCH-2026-10-22.md

Render integration batch — 2026-10-22

Working contract

Personal repository: 0xlillith/solidcraft. Branch: lab/render-scale-resources; draft PR #7, stacked on #5. The fixed upstream snapshot remains unchanged. All checkpoints in this batch are WIP until their consolidated validation and actual application acceptance are recorded.

The owner requested a larger implementation batch and fewer expensive recompilations. Accordingly, source inspection, formatting, JSON validation and prepared tests are distinguished from compiled or executed results. A recovery push is not permission to merge upstream or publish a release.

Integrated inputs

• Renderer checkpoint 5cc4d5fd and progressive-HDR reference correction e7f56218.
• Catalog/SELENE specification branch through 10ada1b4 (PR #3).
• Original PBR/HDR/light authoring resources through 3ffaa5bd (PR #4).
• Native resource-conformance tooling through 6d79d44 (PR #6).
• The resource merge is aafa17f3; both sides' attribution entries and the canonical LF policy were preserved.

The integrated inventory includes 600 material concepts, 280 lighting concepts, 280 environment concepts, 48 authored PBR surfaces with 240 maps, 24 authored HDR fields, 24 lighting rigs and the SELENE demo contracts. Concept definitions and authored bytes are not synonymous with executable runtime features. Original manifests keep their existing readiness labels; runtime adapters must identify their actual supported projection separately.

Implementation slices

Slice	Source status	Acceptance status
Authored branches integrated with ancestry and notices	Done	Source/merge checks only in this branch
Explicit CPU Compact/Studio/Workstation envelopes	Prepared	Consolidated build and runtime validation pending
Profile propagation through saved intent, UI, preflight, capture, BVH, film and host	Prepared	Regression sources added; not yet executed
Searchable catalog, bounded target search and explicit apply commands	Prepared	Source/tests only; no executed application acceptance yet
Authored map/HDR consumers, scalar channels and orientation adapters	Prepared	Native source admission exists; new application/shader tests are not yet executed
Emitter-only/complete rig variants; scene specifications remain gated	Prepared partial projections	Passive cards, target-role scenes and calibration blockers are not silently simulated
Consolidated native/WASM/GUI and representative-scene validation	Deferred until batch freeze	No green aggregate result claimed

CPU profile details and limits are in render-cpu-resource-profiles.md. Old files remain Compact; larger profiles require an explicit choice. Logical quotas are not performance results or promises about available physical RAM.

Preserve throughout the batch

Check and Start must resolve the same request. Save must preserve supported intent without a device. Applying a visual resource must not mutate engineering properties. Missing scene bindings, unsupported effects, calibration blockers and ambiguous light-energy ownership must produce explicit outcomes, not silent approximations. Cancellation, final-only export, source hashes and allocation accounting remain active.

The improved SELENE native source still needs its separate source-lock admission. No obsolete body ordinals or older rover file are substituted for it here. Original authoring-only height maps remain distinct from runtime displacement.

Built-in resource projection checkpoint

The source adds a generated, hash-verified lookup for 192 eight-bit PBR maps and 24 HDRs, a bounded library browser and command API, 96 explicit uniform appearance projections, CPU/GPU embedded HDR setup application, exact GPU emitter conversions and GPU four-channel surface bindings. GPU source/decoded payloads are deduplicated by immutable content plus role across up to 128 bodies. Original authoring readiness flags and height-map exclusions remain unchanged. See built-in library.

Source formatting and generated-lookup checks are not compilation or renderer validation. All new regression and GPU oracle cases are prepared for the consolidated gate. The ongoing old-source gate belongs to the original worktree and is not evidence for this branch. It has reported one GPU protocol-test failure and nine UI failures; those are recorded for explicit reconciliation rather than hidden by claiming a green aggregate run.

Result lifecycle and large-image display

The next source slice pins exactly one displayed completed result independently of a newer completion, while retaining original memory accounting. Tab changes clear device texture caches but preserve each tab's result identity. Save/Reload and Return to Design are outside the long options scroll. CPU saved intent preserves inactive area sampling/seed values while lowering only active area controls into execution; an explicitly disabled denoiser does not select GPU execution.

Large-result display admission uses the captured request's resource profile rather than the current draft or the old preview default. The retained PNG and export stay full resolution. A device/cache-limited display is explicitly labelled and native 100% is disabled for that reduced view; tiled full-resolution viewport inspection is still a future unit.

Prepared regression fixes retain the negative assertions: foreign/unfinished pin requests refuse, old displayed PNG bytes remain exact, tab caches cannot leak, hidden invalid saved values still refuse, and a pending picker still blocks ordinary UI model edits. Static and format checks do not substitute for execution.

Consolidated checks observed on 2026-10-22

The implementation source tree c6793cbfb8af475dcebab9097039544f70ea437c has passed the complete workspace/all-target type check and Clippy with warnings denied. The combined production WGSL parses and validates with Naga (21 entry points), including the new scalar-map reference entry point. This is shader validation, not yet hardware execution evidence for the new channels.

The CPU-feature renderer unit suite completed 451 passed, 0 failed. The original cpu_resource_smoke example then executed the CPU image path at 3840 x 2160, produced a real PNG through cpu-raytrace, verified its dimensions/content, and observed zero retained host bytes after dropping the completed result. SHA-256: 12966e41d9fde0171a14d6380a0b94b5375af8eee207fa34daf581ef2a7dd4f0. A separate synthetic 20,000-body / 40,000-triangle admission was accepted under Studio and rejected under Compact; it was not rendered and is not a performance benchmark. Source identity was checked before/after the test/probe sequence.

The old-source full CI (a0a2b5af..., from the prior worktree) had already recorded GPU/UI failures. Its one remaining engine test process was stopped after verifying its parent chain; it is an interrupted failing run, not a pass and not validation of this branch. The logs and interruption receipt were preserved. This avoided continuing obsolete validation while the corrected batch was being checked.

The new engine, UI and GPU regression sources still require execution on this batch. Application-window acceptance, large complex-scene timing, progressive large-image memory, complete scene presets and portable body-bound image resources remain separate gates. No complete CI pass or release is claimed.

Follow-on portability and staging batch

The previous WASM gate for c0069b6b completed with exit code 0 and unchanged source. Hardware map/HDR probe execution on AMD Radeon 8060S/Vulkan completed three retained PNG images (authored four maps, legacy modulation and scalar control), all with validated PNG framing/hashes and zero retained host bytes after release. This is one representative resource combination, not whole-pack visual acceptance.

Persistent built-in surface bindings are now prepared across the saved setup, engine command path and existing Render library. Four source hashes/version and definition/feature/component/occurrence-path identity are saved; frames use local millimetres and are placed again for every request. A changed, hidden or ambiguous target refuses rather than being rebound to another ordinal. This preserves engineering materials and the CAD timeline. Regression sources cover round trips, undo, separate instances, placement, identity/hash rejection and offline removal. No execution claim is made for these new regression sources until the next gate.

Passive stages and index-matched optical limit

Prepared saved render-only finite ground/cards (32 maximum), original rectangle/AABB clearance, complete numeric rigs with their authored passive cards, explicit stage replacement/removal, and one-time world-mm fitting. Product topology/materials/mass are unchanged. CPU remains direct/specular; adding diffuse cards is not CPU GI. The conservative clearance may reject valid hollow/nonconvex placements.

The combined portability/staging sources passed workspace/all-target Clippy with -D warnings (receipt portability-stage-clippy-r2.json, unchanged fingerprint 6283d730c6eae23364c9b5da965bfb3a35502642e52708a1141eddc15e28e3a7). The first pass found an Arc borrowing type error; it was corrected, not suppressed. This pass predates the following optical fix; final combined validation is pending.

An independently identified issue prevented the staged IOR=1 dielectric from being Lambertian at grazing angles: generic Schlick still approached unit Fresnel. The exact matched-index limit is now prepared consistently in isotropic/anisotropic CPU BRDFs, specular continuations, GPU path and viewport. Conductors retain their weighted grazing response. Matched-index coats contribute no reflection/attenuation. Existing proposal PDFs are unchanged; this is not an importance-sampling speedup. New analytic tests cover grazing directions, metal mixtures and coat identity. See ../render-portable-surfaces-and-stages.md for portable-resource boundaries.

Published content recovery and physical-reference browser

Merged the functional SELENE source branch at cf8e651022deee528b26c303fde58bf3eb8a22a5 and the independently published physical data at 0ddbcf2fbfa62a03bdc945cdbb9bcc6d6e2116ad with their ancestry. The exact unchanged supplied source and functional format-28 master, procedural builders, three native Design previews and retained notices are now available together. SELENE's pinned_native lock supersedes the older awaiting-source state. Its author branch's model/motion results are inherited evidence, not reruns on this renderer. No whole-rover PBR image is claimed here.

Physical resources provide 51 reference profiles, 373 attributed numeric records, 42 sources, 600 concept specifications and 48 surface associations. The extended profiles, qualifiers, conditions, unknowns and omitted native fields are retained. The existing Render library now has a prepared read-only physical category and source/condition inspector. Concepts and surfaces expose their exact candidate, constituent or comparison relations. A new native command prepares, but does not execute, a loss-aware material creation plan after explicit acknowledgement and only with one supported point density. No engineering material or body is assigned. Native card admission and the four authoring-source hashes are checked. The exact resource exports are not regenerated or reinterpreted by this UI consumer.

The hosted resource gate on run 39766733982 failed at attribution on both hosts. Investigation found that an earlier integration had removed baseline inventory rows and licence prose. Restored the exact rows/prose from e7f56218, retaining all new resource rows and adapter notes; no author/licence facts were invented and no gate was weakened. The repository's unmodified xtask::assets::run executed through a small standalone harness and now reports 756 files, all attributed. The 38 physical-tool tests, catalog validator and pinned-native SELENE prepared validator passed locally. Updated hosted results must still be checked separately.

Functional SELENE production-render continuation

The latest published functional master remains cf8e6510 / SHA-256 3bd2bfa1c6336394c4a4208d30b52c80d06658a3bd7ea57cc93b774411d1bd68. Physical resources are integrated through 32eda023: 59 referenced profiles, 46 point-density-eligible projections and 13 missing-density refusals. These references do not silently replace SELENE engineering properties.

Recovered the interrupted replacement-consent correction and retained all three inherited fixture edits. A patch cannot erase an old saved HDR, stage or surface authority to bypass explicit replacement consent.

Commit 9ad37e3b prepares a bounded native GPU envelope, actual-device storage checks, up to 32 disjoint optical bodies, matched output/save/display bounds, and an opt-in reusable headless GPU studio. Whole-workspace/all-target type checking, formatting and authored catalog/source checks passed after the grouped changes. No whole-SELENE final render or complete new CI pass is claimed at this checkpoint. The next execution uses one grouped native CLI build, then reuses the evaluated model and GPU for reviewed requests. See docs/render-gpu-production-envelope.md.

Native master execution and optical review

The fresh native CLI built from 895704a2 (SHA-256 3421d538c571647c1f24fd74d63487c4ba4ab597980cfeea111a01c0db94e699) passed the public-command resource workflow. The renderer crate passed 457 tests / 0 failures. The complete GPU suite first passed 93 tests, with two obsolete limit-rejection fixtures failing and three explicitly ignored tests. edfe6291 reconciles only those tests with the new explicit envelope; it does not change production transport or disable hostile-input checks.

The unchanged functional master evaluated to 2,116 bodies, zero timeline errors/warnings, of which 2,094 bodies / 2,144,099 triangles are visible under the original saved visibility. Actual native GPU drafts include all that geometry, 12 certified glass bodies, six authored PBR surface families and a separate two-triangle passive floor. No CAD geometry was simplified or removed. Five low-resolution camera/lighting trials completed; the selected perspective studio variant preserves the orange/charcoal palette, unclipped gripper and controlled metal highlights. These trials are not the final high-resolution deliverable.

The reproducible selected recipe and independent camera-fit regression tests are in tools/render-library/selene_recipe.py and test_selene_recipe.py. The rig's 1/16 radiance gain and HDR fill are explicit authored adjustments, not a claim that the original calibrated reference ratio is unchanged. The final publication must cite its actual image dimensions, sample/depth controls, full PNG hash and source model hash; no full GUI/CI success is inferred from this native workflow.

Completed full-detail SELENE showcase

The actual native 3840x2400 / 1,024-sample / maximum-depth-16 render completed. PNG SHA-256: 40a492e7979137d0f4027300da3d8af7ca1008c1aab2c309259e65d309a17316; 2,388,114 bytes. The exact immutable-source/renderer/resource/request manifest and the deliberately selected shipped export are assets/render/selene/scenes/studio-hero.json and exports/SELENE-studio-4k.png. Whole-frame and four unscaled detail-crop review accepted the composition and retained detail. Residual native-pixel Monte Carlo grain in some glass regions is explicitly disclosed; this is not a convergence certificate. The loaded native studio was closed and the master hash rechecked unchanged.

The official 12-crate WASM check and 95-test hardware GPU suite are green (three GPU tests explicitly ignored); the render crate has 457 passes. Seven pure Python camera/publication tests pass. Assets and icon validation passed before the final two production exports were added; the final export addition must run attribution validation again.

The focused render-workspace harness initially passed 54 tests with three failures. Investigation identified a real header-layout instability when the conditional dirty label moved the Reload and CPU resource controls between pointer events, plus an obsolete 64-binding test after the explicit 128-binding expansion. The stable-height status row and current-capacity boundary fixture recovered all 57 preexisting workspace tests. The new dedicated hit-target regression is being checked separately against the saved setup's exposure authority; full GUI/CI acceptance is not claimed while that final check is pending. UI-only repairs do not change the native executable that generated the published image.

Closing validation and delivery checkpoint

The focused render-workspace harness now passes 58 tests / 0 failures / 0 ignored, including the new independent hit-target stability regression. The fixed-height dirty-state indication prevents Reload and the CPU resource picker moving between pointer events; the existing clicked save/reload, native CPU configuration, preview/final retention, cross-tab isolation and boundary tests pass. The added fixture uses the saved setup exposure authority rather than an inactive runtime field. This is a harness result, not a fresh interactive GUI acceptance session or whole-project CI approval.

The committed selected showcase passed bounded PNG verification and exact identity checks again; all 760 collected resources are attributed. The package tool copies only the unmodified completed PNG, public manifest, README and checksums. .github/workflows/selene-showcase.yml is restricted to the owner's integration branch and uses the existing pinned checkout/upload actions with read-only contents permissions. It compiles nothing and installs no dependencies. Its artifact status must be reported separately from renderer execution.

Final frozen-source confirmation

On source commit fb0d7a42ac30402deef3e5d408fd7d674f4db414, tree 0bec3bc30e91550709d0494a2f7c5b15a7f57d41, the cached Render-workspace harness also passed 58/58 with hardware-only Vulkan policy, followed by whole-workspace/all-target Clippy with -D warnings and the official 12-crate WASM gate. All three passed and the source-before/source-after check was unchanged. These closing checks do not relink or replace the native executable used for the image.

Read-only artifact workflow 39787390664 completed successfully on that same published commit. Artifact 12428754554 contains only the exact native PNG, public manifest, README and checksums. Archive SHA-256: 74b94e95eb226a65853c424ec6622da641c765df6c63f48cf7765a315aa94d56. Its downloaded archive and every inner checksum were verified independently; the PNG hash remains 40a492e7979137d0f4027300da3d8af7ca1008c1aab2c309259e65d309a17316. Packaging success is distribution evidence, not a new render or aggregate-CI certificate.

The selected high-quality native SELENE image is complete and shipped. The broader renderer roadmap remains open: quantitative convergence/adaptive sampling, further noise reduction on transmitted/reflected paths, EXR/linear-output work, complete scene/advanced-material capability coverage, broader native CPU/GPU acceptance and an interactive GUI/full-project gate are not silently marked done. PR #7 stays draft; no upstream branch, main merge or release was changed.