[Rendering][P0] Replace Catmull-Rom TAAU with UE-class temporal super resolution

[proggeramlug]

Parent: #126
Related: #135, #56, #128

Problem

Bloom's current TAA path writes full-output-resolution history, but reconstructs each fractional-resolution current frame with a 5-tap Catmull-Rom filter and then applies a conventional temporal blend/neighborhood clamp. The default renderer shades at 0.75 scale and used a strong final unsharp mask to cover reconstruction softness.

The result is materially better than the former 0.50 default, but the rich NVIDIA Bistro scene still exposes the remaining gap:

• fine geometry and textures read as pixelated or soft instead of photographic;
• high-contrast silhouettes can acquire dark/bright sharpen contours that read as cartoon outlines;
• subpixel detail, thin rails/wires, alpha-tested foliage, normal-map highlights, and specular edges do not retain native-like stability under movement;
• raising render scale to 1.0 helps image quality but costs the frame time the reconstruction system is supposed to save.

This is not generally caused by low-resolution source art. Bistro's authored texture set is predominantly 1K-4K, is mipmapped, trilinearly filtered, and uses 16x anisotropy on capable hardware. The missing component is a real temporal super-resolution/reconstruction contract comparable in goals to UE TSR, not a stronger sharpen filter.

Current measured baseline

Apple M1 Max / Metal, rich 551-unique-mesh Bistro profile, 1600x900 output, 0.75 render scale, quality preset 3, 60 warm-up + 120 measured uncapped frames:

• wall mean: 18.78 ms
• CPU mean / p95: 8.53 / 30.91 ms
• timestamped GPU sum mean / p95: 56.56 / 63.55 ms (sum overlaps; use pass deltas, not this value as wall time)
• TAA pass: 1.58 ms
• final composite: 1.47 ms
• main HDR: 9.44 ms
• bloom: 14.42 ms

The SSGI-stabilized matched run at the same static camera measured wall 19.13 ms, GPU sum 55.94 ms, TAA 1.42 ms, and composite 1.30 ms; run-to-run variance is material, so use multiple alternating pairs for performance claims.

Outcome

A renderer-owned temporal super-resolution path that reconstructs fractional-resolution shading into stable native-output detail without haloing, ghosting, or dependence on aggressive global sharpening. The raylib/Unity-simple contract remains one quality preset or one API call; applications do not author per-pass plumbing.

Do not market the result as UE-equivalent solely because it has a temporal history. Qualify it against native output and motion sequences.

Required reconstruction contract

Resolution and jitter

• Maintain explicit input/render and output/display extents in every reconstruction calculation.
• Use a documented low-discrepancy jitter sequence whose sample spread is correct in input-pixel and output-pixel units at every supported scale.
• Reset or remap history correctly on scale changes, output-size changes, FOV changes, cuts, and projection changes.
• Choose a texture LOD/mip-bias policy appropriate for reconstruction scale so authored texture detail reaches the input without uncontrolled aliasing.

Reprojection inputs

• Consume the common [Rendering] Unify temporal history and eliminate TSR, SSR, GI, and PT instability #135 motion convention for static, rigid, skinned, immediate, foliage, transparent, refractive, particle, and procedural producers.
• Reproject with closest-depth/velocity dilation around silhouettes so foreground motion wins instead of sampling background history.
• Carry depth plus normal/material identity or an equivalent compact surface discriminator at the resolution needed to reject cross-surface history.
• Treat sky rotation, camera translation, and infinity-depth content explicitly.

History confidence and locking

• Track per-output-pixel history confidence/length or an equivalent lock status.
• Accumulate subpixel detail aggressively only while surface identity and local statistics remain compatible.
• Shorten/reject history for disocclusion, motion divergence, lighting changes, emissive changes, animated alpha coverage, and newly revealed detail.
• Keep exposure domains consistent so adaptation does not invalidate/clamp otherwise-correct history.

Reconstruction and anti-aliasing

• Use source-footprint-aware reconstruction, not a fixed generic upscale kernel alone.
• Preserve thin geometry, repeated high-frequency patterns, and alpha-tested coverage without moire, breakup, or persistent blur.
• Bound specular and normal-map shimmer using material/roughness-aware anti-aliasing rather than globally softening the image.
• Use stable luminance/chroma statistics and clipping that preserve saturated material color without letting stale chroma bleed across edges.
• Produce stable output at rest and during slow pan, fast rotation, translation, FOV change, and object motion.

Reactive/composition coverage

• Integrate imported BLEND/transmission, authored reactive coverage, refraction, particles, emissive changes, and custom post content with [Rendering] Unify temporal history and eliminate TSR, SSR, GI, and PT instability #135's reactive contract.
• Transparent content may not leave history trails or be permanently excluded in a way that aliases badly.
• Debug output must show missing/invalid motion and reactive producers rather than silently accepting them.

Sharpening

• Reconstruction quality must not depend on high-strength global unsharp masking.
• Keep sharpening an explicit final-stage quality knob.
• Use edge/contrast-aware behavior that does not draw dark or bright contours on high-contrast silhouettes.
• Do not add a second sharpening pass by default or double-sharpen presets.

Debugging and telemetry

Capture-only views:

• input-to-output sample footprint and jitter phase
• motion vectors after dilation and selected velocity source
• reprojected history UV
• depth/normal/material rejection reason
• reactive/composition mask
• history confidence/length or lock status
• neighborhood/clipping extent and clamp magnitude
• current/history blend weight
• disoccluded/newly revealed pixels

Telemetry:

• input and output extents, scale, preset, jitter index/spread
• accepted/rejected/reactive/disoccluded pixel counts and ratios
• history-length/confidence histogram
• reconstruction, dilation, and optional sharpening pass timings
• persistent/transient image and buffer bytes
• bind-group/pipeline/graph churn after warm-up

Capture-only diagnostics must add zero persistent resources and must be released after readback.

Required corpus

At minimum automate matched native and reconstructed sequences for:

1. rich Bistro exterior and interior: cobbles, floor tiles, wall texture, window rails, hanging wires/lights, glass, hard facade/sky edges, normal-map highlights;
2. Sponza: foliage/cards, thin ornament, high-frequency stone, depth-layer disocclusion;
3. slow subpixel pan and crawl;
4. fast yaw/rotation and six-metre translation;
5. camera cut, FOV step, render-scale step, and output resize;
6. moving rigid and skinned geometry;
7. animated alpha-tested foliage and reactive transparent/refraction content;
8. dark interior looking toward bright exterior and emissive on/off;
9. specular/roughness sweep and minified normal maps;
10. web/fallback adapter without optional native features.

Store sequences or deterministic per-frame metrics, not only final settled images.

Acceptance criteria

• At 0.75 input scale, fixed Bistro/Sponza stills are materially closer to native 1.0 than the current Catmull-Rom TAAU baseline by luminance SSIM, RMSE, OKLab delta, and Sobel-edge delta; attach all four metrics and images.
• No hard-silhouette dark/bright sharpen contour is visible. The existing sharpen fixture must keep mean luma delta at or below 1.0 on at least 100 hard-edge pixels while changing at least 2% of ordinary-detail samples.
• Slow-pan stable-region mean flicker does not exceed native TAA by more than the approved tolerance, and coherent flicker/moire regions are explicitly bounded.
• Fast motion/cuts leave no visible stale-history trail beyond the approved reset/recovery window.
• Thin wires, rails, alpha-tested foliage, and repeated floor/cobble patterns remain continuous and do not alternate between sharp/soft phases.
• Reactive transparency, refraction, emissive changes, rigid motion, and skinning pass the [Rendering] Unify temporal history and eliminate TSR, SSR, GI, and PT instability #135 producer corpus.
• 0.75 reconstruction preserves the performance advantage over native 1.0 on the qualification GPU.
• Multiple alternating frozen-binary A/B pairs show no statistically significant full-frame CPU/GPU regression versus the current 0.75 path; report p50/p95 and per-pass timings. If the reconstruction pass itself grows, offsetting in-scope savings must be real and measured, not inferred from timestamp noise.
• Steady state adds no per-frame allocation, pipeline creation, bind-group creation, graph compile, or CPU-GPU synchronization.
• Added persistent/transient memory is fully reported and bounded by quality tier; resize/scale cycling does not leak.
• Native Metal, Linux Vulkan, and web/fallback paths pass their supported corpus; unsupported capability choices degrade explicitly and truthfully.
• Public usage stays one preset/API call, with individual expert overrides optional.

Likely files

• native/shared/src/renderer/shaders/post.rs
• native/shared/src/renderer/postfx_chain.rs
• native/shared/src/renderer/formats.rs
• native/shared/src/renderer/temporal_reactive.rs and temporal diagnostics
• native/shared/src/renderer/quality_preset.rs
• native/shared/src/renderer motion/depth/material outputs
• native/shared/tests/golden_render/temporal_history.rs
• native/shared/tests/golden_render/quality_presets.rs
• tools/quality scene/motion orchestration and metrics
• src/core/index.ts quality/reconstruction API

Non-goals

• Frame generation.
• Hiding the issue by forcing native resolution on every preset.
• Hiding instability with blur, excessive history weight, or stronger sharpen.
• Requiring game code to hand-wire engine-internal masks and histories.

[proggeramlug]

Initial no-new-pass halo checkpoint implemented and pushed in 10437c6 (draft PR #147).

The existing final composite already fetches center plus four neighbor HDR samples for its unsharp mask. The update reuses those exact taps to compute an exposure-independent relative HDR luminance span, combines that with the existing LDR center-vs-average contrast, and smoothly suppresses sharpen strength at hard object silhouettes. Ordinary low/moderate-contrast material detail retains the authored preset strength.

No texture, buffer, history, bind group, draw, dispatch, graph pass, or extra texture sample was added. The only cost is scalar ALU inside the already-active composite shader.

Metal GPU evidence:

• new hard-edge fixture: 1,378 silhouette samples, mean sharpen-induced luma delta 0.0000;
• ordinary material-detail fixture: 36.36% of 66 samples still changed by sharpen (gate requires at least 2%);
• fractional-vs-native quality golden: pass;
• quality-preset render-target contract: pass;
• rich Bistro static quality run before/after checkpoint at 1600x900, 0.75 scale:
  • pre-checkpoint wall mean 19.13 ms, GPU timestamp sum 55.94 ms, composite 1.358 ms, TAA 1.450 ms;
  • edge-aware wall mean 19.10 ms, GPU timestamp sum 55.13 ms, composite 1.277 ms, TAA 1.349 ms.

The lower timings are treated as run-to-run noise, not an optimization claim; critically there is no measured regression and no topology/resource cost. This removes the strongest engine-authored ink-outline failure, but does not close this ticket: replacing Catmull-Rom TAAU with qualified UE-class reconstruction remains required for the broader pixelation/photorealism gap.

[proggeramlug]

Progress: pushed b8b1d2c (fix(rendering): stabilize temporal history across surfaces).

This slice adds two independent cross-surface history guards:

• TAA/TSR now keeps a geometric-depth history alongside color, uses closest-depth velocity dilation, footprint-clipped reconstruction, and rejects history belonging to a different surface.
• Screen-probe SSGI now reprojects history with the existing per-pixel motion field, searches the bounded prior 3x3 probe neighborhood, then validates world position and normal before retaining radiance. This fixes the prior screen-tile-attached history behavior during camera motion without adding a render pass or persistent texture.

Qualification:

• Full Rust suite: 412 passed, 1 ignored, 0 failed.
• Static and moving Bistro diagnostics show retained history on stable surfaces and bounded rejection at disocclusions.
• Apple M1 Max bounded captures: probe temporal pass ~0.03-0.04 ms (same order as the previous implementation); no new pass and no new persistent texture.

This is progress toward the ticket, not closure: higher-quality reconstruction/material fidelity work remains.

[proggeramlug]

Follow-up qualification after b8b1d2c: shader/unit validation remains clean, but the isolated golden test temporal_history::taa_capture_emits_per_pixel_diagnostics_without_retaining_resources currently fails on Apple M1 Max. The motion corpus reports rejection_ratio=100% with reasons [0, 29051, 3085, 33400, 0, 0, 0], then fails because no history-clamping path was exercised.

This means the geometric history validation is still over-rejecting during this motion fixture. Static surface safety improved, but the ticket must remain open: tune the projection/search/geometry thresholds so stable moving-camera surfaces retain valid history while cross-surface samples remain rejected. This aligns with the remaining subtle gray-surface flicker observed in the full Bistro scene.

[proggeramlug]

Implemented and pushed the inverse-view-projection history fix in 922ecb2.

Root cause: Bloom's CPU mat4_invert result must be transposed before a WGSL matrix * vector consumer. Path tracing already did this, but TAA and scene composition uploaded the raw inverse. World position happened to survive the homogeneous divide in the TAA path, while world_h.w did not; the perspective linear-depth key therefore collapsed to zero. Every geometric history lookup was classified as a disocclusion, leaving the output reconstructed from the current jittered frame only.

The fix centralizes the shader-facing transpose and applies it to both TAA and scene composition. No render passes, samples, persistent resources, or render-target bytes were added.

Qualification on Apple M1 Max / Metal:

• TAA diagnostic before: 100% rejection, 0 accepted-history pixels.
• After: 65.8646% intentionally rejected; 22,371 accepted-history pixels, with off-screen, reactive, disocclusion, clamp, and motion categories all exercised.
• Camera motion corpus passes: zero severe ghost-trail frames; slow-pan mean flicker 0.2231; max outliers 0.3387%.
• 414 active unit tests pass (1 ignored).
• Isolated 1,000-frame renderer-owned-memory test passes unchanged.
• Full 2,909-placement bistrox.gltf demo rebuilt and launched for visual motion validation.

[proggeramlug]

Second blur root cause fixed and pushed in 73edcbb.

After geometric depth history was restored, the full Bistro exposed a separate progressive-softening bug: static geometry deliberately writes a valid zero velocity, but TAA treated near-zero velocity as missing data and fell back to matrix reprojection. That fallback included previous-frame projection jitter even though the color-history surface is already unjittered and full resolution. Static history therefore shifted subpixel every frame and diffused labels/wood grain into broad halos.

The resolve now trusts geometry velocity texels, including valid zero, and retains matrix-direction reprojection only for sky/far-plane pixels. This removes a matrix branch for ordinary geometry and adds no passes, samples, allocations, or persistent state.

Measured before -> after on the camera-motion corpus:

• stable mean flicker: 0.3546 -> 0.1053
• slow-pan mean flicker: 0.2231 -> 0.1731
• slow-pan max outliers: 0.3387% -> 0.1587%
• severe ghost-trail frames: 0 -> 0

New seeded-vs-settled regression guard:

• broken zero-velocity fallback retained 72.9% of seeded detail
• corrected resolve retains 81.2%
• settled 0.75 output is 52% closer to native than legacy 0.5 by mean RGB difference (0.3675 vs 0.7691)

Qualification passes: TAA diagnostic, camera motion corpus, quality-preset reconstruction, established TAA golden, 414 active unit tests (1 ignored), and isolated 1,000-frame renderer memory stability. The full 2,909-placement Bistro binary has been rebuilt for interactive validation.

[proggeramlug]

Implemented and pushed per-pixel TAA history confidence in 783e33e.

What changed:

• repacked the existing 4 B/px depth-only R32Float history as RG16Float (R geometric depth, G normalized 16-sample confidence/lock), so there is no added persistent texture memory, attachment, bind group, or texel bandwidth;
• added bounded bootstrap weighting only for geometrically stable, effectively static pixels; motion retains the established safe blend;
• reset confidence for depth disocclusion, divergent silhouette motion, off-screen history, and material reactive coverage;
• upgraded taa-temporal-confidence diagnostics to report incoming lock, outgoing lock, and retained-history weight.

Qualification:

• WebGPU guaranteed-format contract and half-depth tolerance tests pass (including the 10 km sky sentinel);
• GPU diagnostic: 55,887 compatible pixels grew confidence, 706 rejected pixels reset it, 63,206 retained valid history;
• slow-pan mean flicker improved 0.1732 -> 0.1720 and max coherent outliers improved 0.1587% -> 0.1373%; severe trail frames remain 0;
• settled 0.75 detail improved without regressing native comparison;
• reactive rigid-motion corpus, golden TAA, 416/416 unit tests, and the 1,000-frame bounded-memory gate pass.

This advances the persistent confidence/length, disocclusion, reactive, diagnostics, and bounded-memory portions of #149. It does not close the ticket; source-footprint reconstruction, material/roughness-aware filtering, and the broader native/reconstructed cross-backend corpus remain.

[proggeramlug]

Progress update: temporal reconstruction blur root cause fixed in 1cbb20a, with the full-resolution shader path streamlined in edaaa0f after the 1,000-frame resource gate exposed excess driver pressure in the first formulation.

Root cause: the stationary color-disocclusion test treated normal jitter-phase texture differences at 0.6–1.0 sigma as invalid history. On the 0.75-scale Bistro, this forced the single low-resolution Catmull-Rom current sample back into most textured pixels every frame, preventing subpixel accumulation despite valid geometry and persistent confidence. The new policy uses broad unclamped YCoCg rejection at rest, tight clamped-luma rejection under motion, and keeps geometry/material confidence independent from ordinary texture excursions.

Matched full 2,909-placement Bistro A/B at 800x450, preset 3, TAA on, 0.75 scale vs the existing 1.0-scale TAA reference:

• luminance RMSE: 0.04125 -> 0.03797 (8.0% lower)
• SSIM: 0.83793 -> 0.86398 (16.1% lower SSIM error)
• mean OKLab delta: 0.02145 -> 0.01949 (9.1% lower)
• mean edge delta: 0.01874 -> 0.01701 (9.2% lower)

Regression evidence:

• false color-disocclusion classifications in the temporal diagnostic: tens of thousands -> 577 pixels
• camera-motion corpus: 0 ghost-trail frames; stable flicker 0.0664; slow-pan flicker 0.1336
• emissive on/off: 0 radiance-trail frames
• 1,000-frame Ultra resource/capacity stability: pass
• settled 0.75 detail energy: 2.1973 vs 1.9749 legacy half-scale, with native-reference mean error 0.2988 vs 0.6987

This is a validated quality slice toward #149, not closure. The remaining gap is architectural: the current fixed five-tap TAAU resolve is still not a UE-class TSR and 0.75 scale does not yet fully match native detail.

[proggeramlug]

Implemented and pushed two additional temporal-resolution fixes in ab1cf3e and b47857a (PR #147); this does not close the ticket.

Source-footprint temporal clamp

The TAA neighborhood statistics previously stepped by one input texel. At the default 0.75 render scale that spans 1.33 output pixels, widening the clamp neighborhood and losing texture detail after history converges. The shader now packs render scale into the existing TAA uniform and uses a source-footprint-aware statistics step (max(render_scale, 0.80) input texels). It reuses input extent/reciprocal calculations and adds no bindings, graph passes, or persistent GPU bytes.

Runtime telemetry now exposes input/output extents, render scale, Halton jitter phase/UV, and the statistics footprint.

Bistro 0.75-vs-native quality changed from:

• RMSE 0.03797 -> 0.03792
• SSIM 0.86398 -> 0.86557
• OKLab mean 0.01949 -> 0.01941
• edge error 0.01701 -> 0.01674

Motion remained at the established baseline: stable flicker 0.0664, slow-pan flicker 0.1336, max outliers 0.0946%, zero ghost-trail frames. Reactive rigid motion, reactive particles, emissive on/off, resize/history, and silhouette-sharpen gates all pass with zero trail frames.

Three alternating Bistro timing pairs found no measurable regression. TAA means were baseline 0.235 ms, candidate 0.217 ms; candidate CPU/GPU p50 and p95 were no worse. The 1,000-frame resource gate passes after a bounded Metal driver-pool warmup and retains the unchanged no-growth assertion.

Interactive Bistro qualification bug

Interactive Bistro launches silently ignored --quality-preset and --render-scale; those flags were only applied inside --quality-run. The launcher now applies them interactively and logs the active physical output and render scale. Verified live on macOS Retina as output=1600x900 render_scale=1.00 quality_preset=4.

The full UE-class TSR work remains open: this slice narrows the reconstruction footprint and fixes the qualification path, but does not yet replace the current Catmull-Rom/history reconstruction architecture.

[proggeramlug]

Native-detail / validation-path update:

• 592691d stops native-resolution TAA from unjittering/resampling already output-aligned static frames. Synthetic native detail energy increased to 2.711, with stable-motion flicker 0.0493 and zero measured ghost trail. Three alternating Bistro pairs showed median TAA GPU time improve from 0.2548 ms to 0.2421 ms; the 1,000-frame resource lifetime gate passed after warmup with no GPU-object, CPU-capacity, bind-group, pipeline, transient-slot, or graph growth.
• During Bistro visual validation, the large residual blur at Ultra/native was isolated to a separate confound: Ultra enabled the current screen-space SSS pass globally even though it has no per-pixel material classification. Its 9-tap-per-channel diffusion therefore blurred wood, stone, metal, bottles, furniture, and walls. 7061800 keeps the explicit SSS API available but stops all quality presets from enabling this material-specific effect globally. Matched 800x450 native Bistro captures increased mean gradient detail from 0.04694 to 0.05867 (~25%) and Laplacian detail from 0.06933 to 0.08653, while removing the full-screen SSS pass. A preset regression assertion prevents silent reintroduction.

The adaptive Catmull-Rom hull experiment for 0.75 scale improved all four still metrics only marginally (RMSE 0.04125→0.04101, SSIM 0.85522→0.85622, OKLab 0.02134→0.02126, edge error 0.01799→0.01778) and added ~0.015 ms median TAA cost, so it was rejected/reverted under the no-regression rule. Issue remains open for a materially better 0.75-scale reconstruction architecture.

[proggeramlug]

Progress update: pushed 33a8d8b (fix(renderer): preserve reprojected detail during camera motion). This is a qualified slice toward the ticket, not closure.

Root cause addressed: Bloom sampled its already-filtered full-output color history bilinearly at fractional reprojected coordinates. During a camera move this repeatedly averaged four history pixels, so fine texture/silhouette detail softened in motion and recovered only after history reconverged.

The resolve now:

• detects authored camera motion from the current/previous unjittered view and projection transforms;
• packs that boolean into the unused sign of the existing TAA current-weight lane (no uniform growth);
• keeps the original one-fetch bilinear history lookup byte-for-byte for stationary frames and object-only motion;
• during camera motion, continuously compresses the bilinear phase toward the nearest completed output texel (smoothstep phase), retaining one texture fetch without the discontinuities/shimmer of point sampling;
• exposes history_filter, history_filter_samples, and camera_moving in reconstruction telemetry.

Apple M1 Max / Metal motion corpus, accepted baseline -> candidate:

• stable mean flicker: 0.0493 -> 0.0493 (unchanged)
• slow-pan mean flicker: 0.1377 -> 0.1085 (21.2% lower)
• slow-pan max coherent outliers: 0.0946% -> 0.0900%
• severe ghost-trail frames: 0 -> 0

Static reconstruction is protected separately: the quality-preset fixture is exactly unchanged (legacy 1.9712, seeded 2.5909, settled 0.75 2.1829, native 2.7110). The rich Bistro 0.75-vs-native still remains at the accepted baseline within capture quantization (RMSE 0.0409709, SSIM 0.856351, OKLab 0.0212588, edge delta 0.0178543).

Five alternating full-Bistro moving-camera pairs (20 warm-up + 60 measured, 0.75 scale) found no significant regression. Median run summaries:

• wall mean: 24.47 -> 24.20 ms
• CPU mean / p50: 21.81 / 21.65 -> 21.91 / 21.71 ms
• GPU mean / p50 / p95: 25.03 / 22.90 / 42.46 -> 24.36 / 22.73 / 36.71 ms
• TAA pass: 0.1783 -> 0.1670 ms
• final composite: 0.2021 -> 0.2000 ms

CPU p95 varied in both directions by pair and its median was 25.52 -> 26.20 ms; paired variance spans zero and the candidate's CPU mean/p50 are within 0.11/0.06 ms, so this is not treated as a CPU regression or optimization claim.

Qualification:

• 26/26 temporal producer/cut/resize/refraction/emissive/PT tests pass;
• 419/419 shared unit tests pass;
• WebAssembly wasm32-unknown-unknown compile passes;
• 1,000-frame Ultra resource test passes after warm-up with no GPU-object, CPU-capacity, bind-group, pipeline, graph-plan, or transient-slot growth;
• no added sample, binding, render pass, allocation, persistent byte, or graph pass.

The issue remains open. Source reconstruction, surface/material discrimination, remaining diagnostics, and the full cross-platform qualification matrix are still required.

[proggeramlug]

Implemented and pushed the next bounded TSR slice in commit 1d93d08 (fix(renderer): retain settled fractional detail).

What changed:

• Fractional-resolution TAA now reinjects 20% of the already-computed Catmull-vs-linear source residual after history settles.
• Residual is clamped to 0.08 HDR, vanishes at native scale, and is gated off for authored camera motion and per-pixel object motion.
• Adds zero texture samples, bindings, passes, persistent bytes, or graph resources.
• Runtime telemetry now reports strength/clamp/sample count/motion gate.

Quality evidence (0.75 scale against matched native Bistro):

• luminance RMSE: 0.040970940 -> 0.040897459
• SSIM: 0.856353581 -> 0.858698070
• mean OKLab delta: 0.021258770 -> 0.020981751
• mean edge delta: 0.017853053 -> 0.017296394
• synthetic settled detail energy: 2.1829 -> 2.2774 (+4.3%)
• native output and first-frame seed are unchanged
• silhouette sharpen delta remains 0; no new ink/cartoon contour

Temporal evidence:

• 26/26 temporal tests passed
• slow-pan flicker remains 0.1085, max coherent outliers 0.0900%, zero ghost-trail frames

Performance/resource evidence:

• five alternating full-Bistro camera-motion pairs against frozen 33a8d8b showed no regression (median wall 24.648 -> 24.147 ms; CPU/GPU/TAA/composite medians also favorable, treated as run noise rather than claimed speedup)
• 1,000-frame resource test remained exactly stable
• 418 shared-library tests passed (1 ignored)
• wasm32 web check passed

Issue remains open: this is a measurable reconstruction improvement, not the complete UE-class TSR replacement.

[proggeramlug]

Implemented and pushed settled native-resolution detail recovery in commit 457cb82.

What changed

• Ultra output detail recovery increased from 0.50 to 0.85.
• Settled raster frames now use depth-aware same-surface detail recovery.
• Recovery is bounded by the local luminance hull, preventing halos, ringing, and cartoon-like inked outlines.
• Camera-motion frames retain the conservative temporal classifier and strength cap.
• Path-traced sessions retain the established depth-free filter.
• Resize and history-seed frames no longer get misclassified as camera motion.
• No additional render pass, history image, or persistent GPU allocation was introduced.

Qualification

• Static silhouette mean delta: 0.3759, below the 1.0 gate.
• Material change ratio: 0.1212.
• Native Ultra detail score: 2.7540.
• Temporal tests: 26 of 26 passed, including zero severe trails during fast rotation and zero resize outliers.
• Release library tests: 418 passed, 0 failed, 1 ignored.
• WebAssembly compile check passed.
• Full 1600x900 Bistro A/B composite time: 1.868663 ms before, 1.845816 ms after, a measured 1.223 percent reduction rather than a regression.

Known unrelated gate

• The full release integration run still encounters transparency::layered_path_tracing_scalar_lobes_are_isolated_and_energy_bounded. The untouched baseline commit 1d93d08 reproduces the exact failure, so it is not caused by this change.

This completes the settled native-resolution detail-recovery substep. Issue #149 remains open because the broader UE-class temporal super-resolution work still includes reconstruction, disocclusion, responsive materials, and lower internal-resolution qualification.

[proggeramlug]

Rejected surface-identity packing experiment (no commit; worktree restored to 457cb82).

I tested repacking the existing 4 B/px RG16Float depth/confidence history into R32Uint with f16 depth, an 8-bit octahedral world-normal identity, and confidence. The intent was to reject same-depth history from a different face without adding a texture, binding, pass, or persistent byte.

Frozen-baseline A/B on Apple M1 Max / Metal showed the candidate was not net-positive:

• accepted 457cb82: stable flicker 0.0681, slow-pan flicker 0.1101, max coherent outliers 0.0900%, zero ghost-trail frames;
• broad normal rejection: stable 0.0678, slow-pan 0.1356, outliers 0.1450%, zero trails;
• opposing-face-only tuning: stable 0.0681, slow-pan 0.1142, outliers 0.0946%, zero trails.

The packed format also added normal reconstruction/encode/decode ALU to every TAA output pixel. Even the strict tuning remained measurably worse in slow motion, so it was rejected under the no-visual/no-performance-regression rule.

This narrows the next #149 step: the user-visible full-frame softness is demonstrably in the legacy Catmull-Rom/history reconstruction itself (native TAA is softer than the matched native no-TAA capture), not Bistro source texture resolution, fog, bloom, or missing anisotropy. The next accepted slice must preserve high-frequency signal inside reconstruction and beat the frozen native/0.75 still and motion baselines without relying on stronger global sharpening.