Hosted Windows compiler evidence
Pre-releaseExplicit hosted Windows compiler evidence — PR #164
Source: f96af2e4217fab8055329e758407495a1ed78ccd, stacked on #163.
PR: #164
The hosted Windows shared lane now explicitly selects DX12/FXC, and the
remaining ordinary GPU test helpers honor backend/compiler environment options.
Production Windows compiler policy, adapter selection, feature requirements,
ignored tests and all rendering assertions remain unchanged. The ray-query
golden helper retains its explicit DXC requirement.
WARP/DXIL failure controls
On unchanged #163 source c643b85, the full library suite crashes with four test
threads when selectable GPU helpers use Microsoft Basic Render Driver. An
eight-thread run passes. Event 1000 identifies d3d10warp.dll 10.0.26100.9278.
The installed debugger catches a second-chance access violation in
d3d10warp!JITBaseVariable::OptimizeCopy, through the WARP DXIL shader-JIT
destructor and compute-pipeline compilation during a traversal submission.
The affected test passes twice alone with DXC and twice with FXC. All ten
hierarchy tests pass serially with DXC, but both four-thread DXC controls crash
within a few seconds. The four-thread FXC hierarchy control passes. Complete
library controls at four and eight threads pass with FXC selected for
environment-aware helpers: 489 passed, one ignored. Some older helpers still
used Auto or a physical adapter during these initial controls; their selection
behavior and commands are preserved, not relabeled as a uniform WARP matrix.
This identifies a failing WARP/DXIL path, not the driver's underlying root cause
or proof that every hosted access violation has the same stack. #164 is a CI
compatibility workaround; the WARP/DXIL defect remains open.
Candidate and hosted verification
The committed Rust/workflow changes exactly match the tested candidate patch.
The candidate's four-thread FXC library run passes 489 tests with one ignored.
Complete physical Radeon DX12/DXC and Vulkan shared components pass, including
all 93 goldens, device negotiation and the remaining integration tests. Existing
capability skips and ignored tests remain recorded. Strict Clippy, formatting,
repository contracts and diff checks pass. DX12/DXC uses a process-local Vulkan
loader override for the separate production device-negotiation helper; ordinary
test helpers honor explicit backend/compiler options directly.
Hosted Tests run 34552411807 passes all 22 jobs. The Windows shared library
stage passes 489 tests with one ignored in 208.25 seconds. Its golden harness
does not admit the CPU adapter, so the runner's nominal 93 golden "ok" results
complete in 0.18 seconds with adapter-skip messages. These are NOT physical GPU
image evidence. The log retains 86 capability/adapter-skip messages across the
job. The physical image evidence comes from the local Radeon runs and hosted
macOS/Metal tests. The native Windows engine build and all 20 example links pass;
the example link phase takes 188.940 seconds. Linking does not prove startup.
The hosted merge checkout has the same Git tree as the exact PR head. Canonical
Metal captures and quality contracts pass. Named RTX 4080 and M1 Max full-quality
jobs remain skipped, and no unavailable hardware result is inferred.
Separate physical FXC failure retained
The expanded physical Radeon FXC suite passes 89 goldens and fails four layered
material tests during pipeline compilation. FXC reports X3570/X3511 for implicit
lookup-table derivatives inside varying light loops. The relevant shader,
specialization, core shader, test definitions and golden helper are byte-identical
to #163. A source-equivalence receipt distinguishes that fact from a separately
executed parent full-suite run. A level-zero LUT correction is being validated
separately; it is not included or claimed as fixed by #164. No tolerance or
baseline change hides these failures.
Archive
Commands, original failure logs, debugger stacks, executable hashes, controls,
complete candidate logs, source receipts, hosted logs and native example build
evidence are included. Every payload is hashed in payload-sha256.json; ZIP CRC
and all payload SHA-256 values are verified. Executables, libraries, memory dumps,
licensed scene assets and junction targets are excluded.
This advances #140 and reports progress on #153/#128. Windows startup, installed
package usability, wider rendering/performance/streaming/API/UI requirements,
physical hardware qualification and draft-stack review remain open. No PR is merged.