Update public inference release, runtime contracts and documentation

[iChubai]

Summary

Update upstream to the public inference release with unified inference/evaluation entrypoints, native model runtimes, scoped diffusion acceleration, strict regression tooling, and bilingual documentation. This is a release-wide source update.

The candidate's tracked files match release commit 62b624f8 exactly, including README. The candidate commit is 4e4c015a; upstream main is included in its history.

User-Visible Behavior

• Run models through typed worldfoundry-eval run commands, the TUI, or Studio; inspect readiness before loading weights.
• Use reversible acceleration with rounding, cache, streaming-state, and checkpoint contracts.
• Compare raw numerical outputs and decoded exported artifacts against accepted references. Missing output, changed shape, non-finite values, and numerical drift fail validation.
• Model pages distinguish runtime evidence, static checkpoint checks, and pending validation; generated launch commands are checked against executable schemas.

Affected Pipeline / Benchmark

• Pipelines: packaged visual generation, navigation/video, 3D/geometry, and embodied-action runtimes under worldfoundry/synthesis and worldfoundry/base_models.
• Benchmarks: packaged evaluation catalogs and runners under worldfoundry/data/benchmarks and worldfoundry/evaluation.
• Runtime profiles: the release inventory under worldfoundry/data/models/runtime and benchmark runtime profiles.
• Public entrypoints: worldfoundry-eval run, zoo, tui, model asset preparation, and Studio inference.

Change Type

• Bug fix
• Model integration
• Benchmark integration
• Pipeline/runtime change
• Documentation only
• Test/QA tooling

Asset / API / GPU Requirements

• CPU contracts use the declared test dependencies; docs require Node 20.19+ and the locked npm dependencies.
• Real inference requires model-specific checkpoints and CUDA environments. The local validation host uses H100 GPUs.
• Hosted providers, gated checkpoints, datasets, and simulators require their declared credentials or assets when selected.

Commands Run

The passing source checks below correspond to the release tree, which is identical to this candidate.

Command	Result	Notes
make test PYTHON=python3	Pass	805 tests, zero skips; complete selection and source checks.
make test-docs-contracts PYTHON=python3	Pass	29 tests, zero skips; published commands and catalog claims.
make docs-check PYTHON=python3	Pass	Generated API, model, benchmark, and coverage data.
bash scripts/docs/build.sh --skip-bootstrap	Pass	Type checking and static site build.
make lint PYTHON=python3	Pass	Formatting, source, manifests, and registries.
python3 tests/manual/inference_regression_coverage.py --base origin/main --output /tmp/worldfoundry-release-certification-20261005/upstream-coverage.json	Fail	The full upstream comparison affects 238 native variants; 218 lack a replay recipe.

Release CI passed geometry impact, inference replay, CPU tests, inference tensors, public surface, and packaging/license checks for 62b624f8.

Checkpoint / API Key Needs

• Checkpoints and weights: per-model manifests; staged outside the source tree.
• Environment variables: model asset roots and provider credentials declared by the selected runtime. Replay suites use WORLDFOUNDRY_REGRESSION_* fixture settings.
• API providers: optional, selected through runtime profiles.
• Local cache: existing checkpoints and fixture inputs are required for real GPU replay.

Sample Artifact Evidence

• Public source evidence: release CI results and artifacts.
• The native replay matrix contains 42 cases, including camera-conditioned DreamX, two-chunk spatial Alaya, and conditioned DIAMOND CS:GO.
• Candidate GPU certification and complete coverage of the upstream comparison remain pending. Real inference evidence is tied to its executed source revision, inputs, weights, and environment.

Validation Matrix Status

Area	Status	Evidence / Command
Unit or focused regression	Pass	Public CPU and documentation contracts; identical release tree.
Real inference validation	Not run for candidate	Exact-commit GPU certification pending.
Streaming or multi-turn path	Pass for CPU contracts	Release inference tensor checks; candidate checkpoint trajectories pending.
Benchmark runner / metric path	Not run at full benchmark scale	Declared assets and official runtime requirements apply.
Docs build or link check	Pass	Docs build and generated-data checks.
GPU/API-key dependent path	Not run for candidate	Requires model-specific assets and environments.

Leaderboard Validity Impact

• Readiness remains scoped to the evidence recorded for each model and benchmark.
• Demo, static, and partial validation evidence does not establish full benchmark or leaderboard validity.
• Source tests, preflight declarations, and numerical parity are separate from official benchmark scoring.

Compatibility And Risk

• The public source layout and entrypoints are updated across the release; downstream users should follow the maintained CLI and model pages.
• Complete GPU replay requires local assets and model-specific environments.
• The upstream comparison has 218 native variants without replay recipes; complete candidate GPU certification remains pending.

Checklist

• I kept the change narrowly scoped. This is a complete release update.
• I did not commit secrets, API keys, large checkpoints, or generated cache files.
• I reused packaged fixtures where practical.
• I updated docs for public behavior, installation, and entry commands.
• I documented external asset and runtime requirements.
• I kept readiness and leaderboard claims scoped to their evidence.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.