Stability benchmark paths

bistro_exterior_stability.camera.json and bistro_interior_stability.camera.json are deterministic 420-frame paths for the default 120-frame warmup and 300-frame measurement run. Both include a static interval, slow pan/orbit/dolly, fast strafe/turn, stop-and-go recovery, yaw wrapping, and a camera cut. The matching project presets are:

• projects/benchmark_interactive.lookdevpt.json and projects/benchmark_reference.lookdevpt.json for Bistro Exterior;
• projects/benchmark_interactive_interior.lookdevpt.json and projects/benchmark_reference_interior.lookdevpt.json for Bistro Interior.

bistro_exterior_static.camera.json holds the Exterior camera for all 420 frames. It is intended for temporal-CV captures after a long static warmup, so stop-and-go recovery is not mixed into the stationary-noise measurement.

Example:

Bin\x64\Release\D3D12LookDevPTWinUI.exe `
  --project projects\benchmark_interactive.lookdevpt.json `
  --benchmark `
  --benchmark-kind performance `
  --camera-path benchmarks\bistro_exterior_stability.camera.json `
  --frames 300 `
  --warmup 120 `
  --seed 1 `
  --output benchmark-output\bistro-exterior

Camera-path angles are radians. A keyframe with cut: true holds the preceding pose until that exact frame and then changes pose without interpolation. Yaw is interpolated over its shortest wrapped arc. Benchmark time is fixed at 1/60 s.

--benchmark-kind accepts three values:

• performance disables the full-resolution quality-counter pass and is the only kind eligible for the 60 fps gate;
• quality enables history-rejection, contribution-energy, and finite-value diagnostics plus capture-friendly explicit intermediates;
• combined preserves the original all-in-one behavior and remains the default when the option is omitted.

The output directory contains per-frame frames.csv, aggregate summary.json, the final HDR/LDR/AOV captures, artifacts.json, and quality_analysis.json. The artifact manifest records each capture's role, dimensions, source format, byte size, SHA-256 digest, channel min/max, and NaN/Inf counts. summary.json.performanceGate is evaluated only for an isolated performance run at native 1920x1080 with at least 120 warmup and 300 measured frames whose delayed GPU timestamps match every submitted frame. Short smoke runs still write all diagnostics, but report eligible: false and cannot pass the gate.

The performance thresholds are GPU p95 <= 16.7 ms, GPU p99 <= 20 ms, CPU p95 <= 4 ms (GPU fence throttling is reported separately), and frame/history resources <= 512 MiB. Eligibility also requires an RTX 4070, a Bistro target scene, interactive_game, Beauty view, the prescribed ray budget, active NRD REBLUR and active RTXDI GI+DI without fallback. A Baseline fallback is intentionally ineligible rather than being reported as a passing final-pipeline result. For a release-quality result, run the same path three times and use the median run when comparing builds.

Measured-frame sequences

The original CLI remains unchanged and captures only the final artifact set. Add --capture-every N to also capture LDR and HDR beauty for every Nth measured frame (warmup is never captured). Add --capture-aovs to include the SurfaceGuides and lighting signals in each scheduled frame:

Bin\x64\Release\D3D12LookDevPTWinUI.exe `
  --project projects\benchmark_interactive.lookdevpt.json `
  --benchmark `
  --benchmark-kind quality `
  --camera-path benchmarks\bistro_exterior_stability.camera.json `
  --frames 300 --warmup 120 --seed 1 `
  --capture-every 1 --capture-aovs `
  --output benchmark-output\bistro-exterior-sequence

Per-frame files use frames/<measured-frame>/.... Capturing every frame is intentionally explicit because native-resolution HDR plus AOV sequences can be very large and each readback blocks to preserve exact frame association.

quality_analysis.json selects the contiguous beauty_hdr inputs for a post-process temporal luminance CV calculation and defines its median 1% / p95 3% targets. It also contains the 10–90% edge-width contract; copy it into an analysis request and supply a high-SPP reference manifest plus an edge ROI. GPU history-rejection and contribution-energy counters are versioned in summary.json.metricSchema.qualityCounters. Until a backend publishes a counter, it is explicitly marked available: false rather than recorded as zero.

Run the harness unit test with:

Scripts\TestBenchmarkHarness.ps1

Scripts/RunStabilityBenchmarks.ps1 automates the three-run Exterior/Interior suite and writes suite-summary.json with the median-p95 run selected for each target. Use -BenchmarkKind performance|quality|combined to select the run contract. Add -IncludeReference to capture the matching unbiased Baseline PT profiles alongside Interactive. -CaptureEvery 1 -CaptureAovs enables the full measured-frame HDR/LDR/AOV sequence for temporal-quality analysis. -BackendMatrix generates and runs all four render modes against NRD REBLUR, NRD RELAX, the internal fallback, and OFF without modifying checked-in presets. Scripts/BuildBackendMatrix.ps1 rebuilds and smoke-launches the optional-NVIDIA feature matrix, including individual and combined EnableNRD=false, EnableRTXDI=false, and EnableDLSS=false configurations.

After a sequence capture, Scripts/AnalyzeBenchmarkSequence.py <output-dir> --start <static-segment> --count 32 --enforce computes the median/p95 linear HDR luminance temporal CV and scans every sample for NaN/Inf. The surface CV uses primaryHitT from the alpha channel of surface_motion_hit.dds (0 = miss, > 0 = hit): complete per-frame AOVs are intersected across the analyzed frames, otherwise the final AOV is used as a static-camera mask. A partial per-frame AOV set is rejected instead of being silently mixed with the final mask. The output records the mask source, artifact count, and surface-pixel count. Zero-mean surface pixels have undefined CV and are reported separately; no luminance floor is applied. It requires NumPy and writes temporal-analysis.json with the unchanged 1%/3% acceptance gate. Supplying a matched --reference-hdr also evaluates the strong-edge 10-90% width ratio against the 1.15 blur limit.

The HDR reader accepts DirectXTex's adaptive output, where an individual scanline can fall back from channel-RLE to raw RGBE when compression would be larger. The sequence may legally mix both row encodings in one file.

The DDS parser and mask/CV behavior can be tested with:

Scripts\TestBenchmarkSequenceAnalyzer.ps1 -PythonExecutable <python-with-numpy>