Fix FXC layered-material compilation with explicit lookup LOD

docs/evidence/windows-fxc-layered-lut-v1.md

@@ -0,0 +1,57 @@
+# FXC layered-material lookup sampling
+
+The expanded physical Radeon DX12/FXC run on #164 passes 89 rendering goldens
+and fails four during `scene_layered_pbr_pipeline` compilation:
+
+- clearcoat-normal minification and motion;
+- explicit mirrored-tangent anisotropy;
+- anisotropy under negative model scale;
+- layered materials across opaque, sorted, reactive and weighted paths.
+
+FXC reports X3570 for a gradient instruction in a varying loop, followed by
+X3511 when it cannot unroll the loop. The shader, scene specialization, core
+shader, test definitions and golden helper are byte-identical between #163 and
+#164. The source-equivalence receipt distinguishes that comparison from a
+separately executed parent full-suite run. These are compiler failures before
+image assertions, and their original logs remain in the #164 evidence.
+
+## Correction
+
+`layered_sheen_directional_albedo` used an implicit-derivative texture sample.
+The helper is called from varying direct-light loops. Its R16Float lookup table
+has exactly one mip, so explicitly sample level zero with the same texture,
+sampler and lookup coordinates. This removes the implicit gradient requirement
+that forced FXC to unroll the surrounding loop. No additional binding, texture,
+allocation, feature requirement or compiler fallback is introduced.
+
+The change applies to the shared layered-material helper, including its scene,
+transparency and visibility uses. The existing sampler policy remains intact;
+the successful image assertions below are the evidence for rendering behavior,
+not an assumption that compiler outputs are byte-identical.
+
+## Validation
+
+The complete local DX12/FXC shared component passes: 489 library tests, device
+negotiation, all 93 goldens and the remaining integration checks. All four
+previously failing tests now create their pipelines and pass their original
+rendering assertions. Existing one ignored library test, four ignored goldens
+and two ignored documentation examples remain unchanged.
+
+Raster goldens require and identify the Radeon 760M/DX12 adapter. Ordinary
+pipelines use explicit FXC; the ray-query golden helper deliberately requires
+DXC. Older library helpers can explicitly select software adapters. The full
+FXC component takes 1,073.406 seconds including compilation, shader creation and
+tests; this is test duration, not an FPS or GPU-budget measurement.
+
+DX12/DXC and Vulkan each pass all 93 goldens with four existing ignored tests.
+Strict Clippy, formatting and repository contracts also pass. The test driver
+records the exact candidate shader patch against `f96af2e`; only report updates
+followed these runs. Hosted validation is pending.
+The WARP/DXIL concurrent compiler crash documented with #164 remains a separate
+open defect. This change does not qualify unavailable named hardware or wider
+performance, starter, packaging or engine-plan acceptance.
+
+Source patches, the original failure log and source-equivalence receipts,
+commands and complete results are retained under
+`tools/quality/out/windows-engine-plan/fxc-layered-lut/` and
+`tools/quality/out/windows-engine-plan/windows-shared-crash/`.

docs/evidence/windows-warp-compiler-v1.md

@@ -37,8 +37,16 @@ The ray-query golden helper retains its deliberate DXC requirement.
 This is a software-renderer CI workaround. The WARP/DXIL access violation remains
 open, and physical DX12/DXC qualification remains required. The ordinary Windows
 engine compiler policy is unchanged. The candidate's library suite passes with
-four threads and explicit FXC (489 passed, one ignored). Complete local shared
-components and hosted validation are still running.
+four threads and explicit FXC (489 passed, one ignored). Complete local
+DX12/DXC and Vulkan shared components pass, including all 93 goldens.
+
+At source `f96af2e`, hosted Tests run 34552411807 passes all 22 jobs, including
+actual Windows shared execution, the native engine build and all 20 example
+links. Hosted Windows goldens skip its CPU adapter and do not qualify physical
+images. The [published report](https://github.com/Bloom-Engine/engine/releases/tag/quality-evidence-warp-compiler-20260911)
+retains the skips and all original crashes. The expanded physical FXC run also
+exposes four [layered-material compiler failures](windows-fxc-layered-lut-v1.md);
+that separate correction preserves the original rendering assertions.
 
 Diagnostic commands, logs, debugger stacks, executable hashes and results are
 retained in `tools/quality/out/windows-engine-plan/windows-shared-crash/`.

docs/windows-engine-plan.md

@@ -17,9 +17,9 @@ first nine-scene Radeon evidence are in draft PR #154. Follow-up work starts at
 | #127 Vulkan PT correctness | Three deterministic progressive and motion runs, both negative controls, finite intermediates, reset/lighting/rigid-motion checks, retained report | Canonical hardware gate, all four focused temporal tests, and CPU reference sanity check pass on Radeon/Vulkan; [report](evidence/issue-127-windows-vulkan-v1.md) and [raw evidence](https://github.com/Bloom-Engine/engine/releases/tag/quality-evidence-155-windows-vulkan-20260910) published |
 | #128 Windows image discrepancies | Identify the first incorrect stage or document a reviewed backend-specific baseline decision; rerun the full strict corpus and reproducibility checks | Cutout and surface corrections pass all nine Radeon images. At #159 source `d610d6a`, full runs 2 and 3 pass every configured check and reproduce 257 artifacts byte-identically with matching metadata and timing differences inside existing noise bounds. Earlier invalid runs retain their failures; named hardware acceptance remains separate |
 | #135 / #149 temporal reconstruction | Enforced motion/producer/quality-preset corpus, representative scenes, fractional/native and frozen A/B timing, memory/resize checks, platform evidence | Device/resource, stationary SSGI, and profiler fixes are retained. The surface correction passes original HD startup limits and 154,720 analytic receiver checks on Vulkan, DX12, and hosted Metal; 93 local goldens pass, including lighting recovery. The full Radeon corpus passes twice. Wider representative scenes, frozen A/B performance, memory/resize, and platform acceptance remain open |
-| #140 integration gates | Same required local/hosted lanes pass on exact source; release package startup and all-example evidence | #160 fixes silent Windows CI non-execution and MSVC PATH ordering. #161 passes the actual native engine build and all 20 native links locally and in hosted CI; [evidence](https://github.com/Bloom-Engine/engine/releases/tag/quality-evidence-windows-examples-20260911) is published. Draft #162 corrects the two focused DX12 GPU failures and passes the complete Vulkan shared component. Hosted Windows still crashes with an access violation; the expanded local DX12 goldens expose a separate camera-history reset defect. Release startup/install acceptance remains open |
+| #140 integration gates | Same required local/hosted lanes pass on exact source; release package startup and all-example evidence | #160 fixes silent Windows CI non-execution and MSVC PATH ordering. #161 passes the actual native engine build and all 20 native links locally and in hosted CI. #162 fixes the focused DX12 failures; #163 fixes camera-history reset. #164 passes all 22 hosted Tests jobs using an explicit FXC Windows lane. The underlying WARP/DXIL crash remains open. A separate layered-material correction passes the full local FXC shared suite and all 93 DXC/Vulkan goldens. Release startup/install acceptance remains open |
 | #138 capability fallback | Actual constrained-adapter startup and relevant forced-tier corpus, truthful capability outputs | Existing implementation/evidence preserved; physical constrained-limit acceptance still needs proof |
-| PR integration | Reviewable changes, passing required checks, full issue evidence, merge-ready rendering branch | #147 and the stacked fixes #154–#159 remain drafts; no merge performed |
+| PR integration | Reviewable changes, passing required checks, full issue evidence, merge-ready rendering branch | #147 and the stacked fixes #154–#164 remain drafts; no merge performed |
 
 ## Engine work retained in scope
 
@@ -67,14 +67,26 @@ audit are saved in `tools/quality/out/windows-engine-plan/plan-requirements.json
    goldens and the expanded cut check; its [evidence is published](https://github.com/Bloom-Engine/engine/releases/tag/quality-evidence-camera-history-20260911).
    The Windows library crash reproduces locally inside WARP's DXIL shader
    optimizer during concurrent traversal tests. An [explicit FXC hosted lane](evidence/windows-warp-compiler-v1.md)
-   and consistent test compiler selection are being validated. The underlying
-   WARP/DXIL defect and physical DX12/DXC acceptance remain separate.
+   and consistent test compiler selection pass all 22 hosted Tests jobs at #164,
+   with [published evidence](https://github.com/Bloom-Engine/engine/releases/tag/quality-evidence-warp-compiler-20260911).
+   Hosted Windows goldens skip its CPU adapter; complete physical Radeon DX12/DXC
+   and Vulkan shared runs pass all 93 goldens. The expanded physical FXC run
+   exposed four layered-material compiler failures. A [level-zero LUT correction](evidence/windows-fxc-layered-lut-v1.md)
+   restores the complete local FXC shared component, including all 93 goldens;
+   all 93 DXC and Vulkan regression goldens also pass. Hosted checks for that
+   correction are pending. The underlying WARP/DXIL
+   defect remains open.
 2. Finish all-example native linking, real starter/example startup, and clean
    Windows installation. The [native example gate](evidence/windows-example-gate-v1.md)
    passes all 20 links locally using Perry 0.5.1220 and one matching source-built
    runtime profile. It adds required full-lane Windows PR compilation and
    rejects missing or stale executable outputs. Hosted example validation also passes at `59244b9`; actual startup and clean
-   package installation remain required.
+   package installation remain required. Packing and installing the actual
+   package succeeds in a clean project, but its installed `bloom-web --help`
+   command fails on Windows because npm's shim tries `/bin/bash`. A bounded
+   starter compiles to WASM from that installed package; runtime rendering is
+   not yet proven. The next implementation work is portable command execution
+   and native/browser startup acceptance.
 3. Complete the wider temporal/geometry, performance, memory, resize, and
    capability corpus. The
    [HD surface correction](evidence/windows-ssgi-surface-v1.md) and two valid

native/shared/shaders/layered_pbr_v3.wgsl

@@ -403,10 +403,14 @@ fn layered_sheen_visibility(
 }
 
 fn layered_sheen_directional_albedo(n_dot: f32, roughness: f32) -> f32 {
-    return textureSample(
+    // This lookup table has one mip. Direct-light loops can have varying
+    // iteration counts, so avoid the implicit derivatives that force FXC to
+    // unroll those loops when sampling it.
+    return textureSampleLevel(
         layered_sheen_albedo_tex,
         layered_sampler,
         vec2<f32>(clamp(n_dot, 0.0, 1.0), clamp(roughness, 0.0, 1.0)),
+        0.0,
     ).r;
 }