feat: temporal guides + still-scene convergence fix

[DennisSmolek]

Temporal guides program (M1–M6) + the post-parity core work, plus a still-scene
convergence fix found by the program's first full-pipeline consumer.

Merging this triggers publish.yml path B — the branch carries feat: commits
since v0.1.0, so it auto-bumps and publishes 0.2.0.

Temporal guides — a new public surface

Dilated motion, dilated depth, and disocclusion are frame properties, not
upscaler properties: every temporal effect upstream (SSGI/SSR reprojection,
SVGF-class denoisers, TAA) re-derives worse versions privately. The upscaler now
publishes its internal working set as upscaler.guides (ordinary three
textures), and the frame can be driven split so the geometry guides exist
before the final color does:

upscaler.dispatchGuides({ depth, velocity }, camera); // right after the G-buffer
// … effects sample guides.dilatedMotion / .disocclusion / .dilatedDepth …
upscaler.dispatchUpscale({ color, deltaTime }, camera);

• path: 'guides' runs the early stage alone, for apps that never upscale.
• Reactivity is bidirectional — an explicit mask merges (per-pixel max)
  with the auto-generated one; effects can write into guides.reactive mid-frame.
• MomentsPass — standalone, signal-agnostic (E[x], E[x²]) over any float
  texture plus one coarse level, with zero coupling to the upscaling pipeline.
• temporalGuides(depth, velocity, camera) publishes the same bundle as TSL
  texture nodes; upscale(..., { guides }) shares one computation across the graph.
• Contracts per product (format, space, resolution, latency) live on the
  TemporalGuides type and in TEMPORAL-GUIDES-SPEC.md. Live references:
  examples/12-temporal-guides (raw), examples/13-guides-node (TSL).

Cross-repo acceptance (M6): PASS. An external SSGI/SVGF consumer ran its
identical temporal stack against its private guides pass vs this bundle:
still-camera stability bit-identical, teleport reconvergence 1.275 s vs
1.288 s, across-arm diff below within-arm noise. Their verdict — drop-in
replacement. @experimental is therefore off the raw surface and MomentsPass;
it stays on the TSL node only, whose contract is the accepted one but whose
graph plumbing has no external consumer yet.

The monolithic dispatch() is unchanged in behavior: byte-identical captures
and perf-neutral (−2.7%, within noise) against the pre-split pipeline.

Core fixes

Still-scene convergence (consumer report 3, confirmed ours). On a still
camera the output never converged — flickering Disocclusion silhouettes and a
forever-rolling AccumulationAge, at every ratio including NativeAA. Reproduced
in our own bench with no consumer code. The decisive metric was the frame diff
between two frames at the same jitter phase one period apart, where benign
per-phase shimmer nets ~0 and we measured 0.182. Three stacked defects:

1. Depth-clip vote starvation (reconstruct.ts) — only taps in front of the
   current surface could vote or carry weight, so one bilinear tap straddling
   last frame's texel-quantized dilated-depth boundary became the sole voter and
   disoccluded still edges every phase. Now every valid tap votes and the best
   tap wins; genuine trails still read ~1.
2. Clip-magnitude history aging (accumulate.ts, removed) — aging the sample
   count by clip strength pinned equilibrium age low at any contrasty edge, so
   alpha never shrank. FSR2 never does this.
3. Clip write-back (accumulate.ts) — the blend stores the clipped
   history, so each phase's variance box re-snapped the buffer regardless of
   alpha. Fixed with STILL_CLAMP_RELAX, gated on stillness × convergence ×
   absence of disocclusion/shading-change/reactivity.

Q1 2×: 0.211 → 0.112 consecutive, 0.182 → 0.018 phase-locked. NativeAA:
0.081 / 0.003. New Q12 cornell scenario: 0.024, disocclusion view black, age
saturated. Motion scenarios (Q3/Q4) verified regression-free.

Grazing-angle disocclusion flicker — jitter-delta-compensated reprojection
plus a tolerance widened by the dilation ring's own depth relief.

Post-parity adoptions (all measured)

• RCAS sharpens in conditioned tonemap space, inverting once — −34% RCAS,
  −5.7% total, capture-identical.
• Host pre-exposure (preExposureTexture) honored end-to-end: DeltaPreExposure
  history correction + host-invariant metering. Byte-identical when absent.
• AMD's viewport/depth-scaled disocclusion tolerance, inside our fused pass.
• Multi-scale shading-change detector (shadingChange.ts) replacing the 3×3
  heuristic — 0.044 ms, 5× cheaper than the source-style candidate, measurably
  fewer false positives under motion.

Tooling

• scripts/measure-convergence.mjs — headless CDP convergence meter
  (consecutive + phase-locked frame diffs, debug-view PNGs).
• Bench scenarios Q11 (host pre-exposure) and Q12 (cornell still-convergence).
• Parity candidate bundles remain runnable for A/B.

Verification

npm test (172 passed) · npm run typecheck · npm run lint · npm run build
all clean. GPU verification is recorded per item in bench/docs/NEXT-STEPS.md
and TEMPORAL-GUIDES-SPEC.md; consumer-side evidence in
GUIDES-HANDOFF-RESPONSE.md.

Known follow-up (not blocking)

src/shaders/candidate*.ts — ~55 kB of parity-research shader source — is
statically imported by Upscaler.ts and so ships in dist. Bench-only, gated
behind _candidateBundle, and not tree-shakeable as wired. Worth moving behind
a bench-only entry point in a later release; deliberately not touched here,
since it means editing a core path right before a publish.

🤖 Generated with Claude Code

[DennisSmolek]

Follow-ups from this branch are now tracked as issues rather than left in the description:

• Keep parity-candidate shader bundles out of the published build #2 — keep the parity-candidate shader bundles out of the published build (the known follow-up noted above)
• Drop @experimental from the TSL guides node once an external consumer wires it #3 — drop @experimental from the TSL guides node once an external consumer wires it
• Cross-device verification: every measurement is from one Apple Metal adapter family #4 — cross-device verification (every measurement here is from one Apple Metal adapter family)
• STILL_CLAMP_RELAX: measure the ghosting tradeoff under sub-detector lighting drift #5 — measure the STILL_CLAMP_RELAX ghosting tradeoff under sub-detector lighting drift
• Bench capture scenario for merged reactive masks #9 — bench capture scenario for merged reactive masks (deferred at guides M3)
• Publication-grade evidence gaps in PAPER-NOTES #10 — publication-grade evidence gaps in PAPER-NOTES
• Auto-release publishes to npm but creates no GitHub Release or changelog #11 — auto-release publishes to npm but creates no GitHub Release or changelog

None block merging.