iChubai · Oct 20, 2026
diff --git a/README.md b/README.md
@@ -114,17 +114,15 @@ or [中文](https://openenvision.github.io/WorldFoundry/zh/docs).
 | Use the browser workspace | [Studio](https://openenvision.github.io/WorldFoundry/docs/guides/studio) |
 | Tune inference | [Inference optimization](https://openenvision.github.io/WorldFoundry/docs/guides/inference-optimization) |
 | Score generated outputs | [Evaluation](https://openenvision.github.io/WorldFoundry/docs/evaluation) |
-| Validate changes and understand internals | [Validation](https://openenvision.github.io/WorldFoundry/docs/reference/validation) · [Architecture](https://openenvision.github.io/WorldFoundry/docs/maintainers/architecture) |
 
 ## Contributing
 
 See [CONTRIBUTING.md](CONTRIBUTING.md) for GPU development setup, coding conventions,
 tests, DCO sign-off, and the PR workflow.
 
-Repository [Agent Skills](skills/index.md) provide opt-in workflows for model and
-benchmark integration, inference optimization, numerical validation, and maintenance.
-
-Use these checks before opening a PR or marking a model/benchmark ready:
+- [Architecture](https://openenvision.github.io/WorldFoundry/docs/maintainers/architecture) — module boundaries and execution flow.
+- [Validation](https://openenvision.github.io/WorldFoundry/docs/reference/validation) — fast checks, CUDA contracts, and real-checkpoint evidence.
+- [Agent Skills](skills/index.md) — model integration, optimization, and maintenance workflows.
 
 Community standards and private vulnerability reporting are covered by the
 [Code of Conduct](CODE_OF_CONDUCT.md) and [Security Policy](SECURITY.md).
@@ -134,11 +132,6 @@ Community standards and private vulnerability reporting are covered by the
 If you use WorldFoundry in research, cite this repository and the upstream methods,
 checkpoints, datasets, and benchmarks that your work depends on.
 
-## Community Policies
-
-Contributions follow the [Code of Conduct](CODE_OF_CONDUCT.md). Use the
-[Security Policy](SECURITY.md) for private vulnerability reports.
-
 ## License
 
 WorldFoundry-authored code is licensed under the [Apache License 2.0](LICENSE), except where a file or component carries a different license notice. Vendored and adapted upstream components remain subject to their original licenses, collected in [THIRD-PARTY-NOTICES](THIRD-PARTY-NOTICES).

diff --git a/docs/fumadocs/content/docs/reference/cli.mdx b/docs/fumadocs/content/docs/reference/cli.mdx
@@ -67,6 +67,15 @@ Running `worldfoundry-eval` with no arguments prints a short first-run banner. A
 
 **Which command.** Everyday reproducible / model-page commands use `worldfoundry-eval run <id> --pipeline.task-profile …`. Use `python -m worldfoundry.studio.workspace_job infer` only for Workspace visual smoke.
 
+Within one Workspace process, `WORLDFOUNDRY_STUDIO_BATCH_MAX_SIZE` (default `1`)
+and `WORLDFOUNDRY_STUDIO_BATCH_DELAY_MS` (default `5`) configure compatible native
+Wan T2V and 5B TI2V resident-worker batches. Admission requires parallel job
+slots, resident weights and compatible Torch attention; compiled batches use
+regional scope. Configuration takes effect before worker startup and does not
+combine independent CLI processes. See the
+[batch admission, warmup and cancellation contracts](/docs/guides/inference-optimization#studio-resident-queue-batching)
+for precision and input requirements.
+
 Examples below use `worldfoundry-eval` so shell history, job scripts, and bug reports share one spelling.
 
 ## Help at every level
@@ -340,6 +349,16 @@ Audio uses an `input_audio` block with base64 `data` and `format` set to `wav` o
 other audio formats. Text and image blocks retain their content and data URIs.
 Conversion is local and does not call an API provider.
 
+### Job completion and cancellation
+
+Durable command jobs preserve the original termination reason in completion
+receipts and restored status. Timeouts remain `failed` with their timeout error;
+user cancellations remain `cancelled`.
+
+On POSIX, bounded benchmark commands clean up their owned process group before
+returning or propagating an interruption. Timeout payloads record `timed_out`
+and `kill_stuck`; the latter indicates that teardown did not finish.
+
 The parser is the source of truth for flags. Use `worldfoundry-eval <command> [<subcommand>] --help` instead of copying flags from an unrelated model or benchmark recipe.
 
 ## Common workflows
@@ -483,6 +502,28 @@ Benchmark-specific inputs such as data roots, score directories, prompt manifest
 
 For HarnessEval-W, `--result-model-id` selects a model without changing the official common-case cohort. See [HarnessEval-W](/docs/evaluation/benchmark-hub/harnesseval-w) for Python execution, cached scoring and model prerequisites.
 
+### 6. Replay native Wan execution and request batches
+
+From a checkout with local checkpoints:
+
+```bash
+python -m benchmarks.inference.wan_execution_preparation \
+  --checkpoint /path/to/Wan2.1-T2V-1.3B \
+  --output-dir tmp/wan-execution --request-batching
+```
+
+`--request-batching` adds A/B, B/A, A/B batches to the singleton comparison.
+Select `wan2.1-t2v-1.3b`, `wan2.2-t2v-a14b`, `wan2.2-ti2v-5b` or
+`wan2.2-i2v-a14b` with `--model-id`. Image-conditioned replays require
+`--images IMAGE_A IMAGE_B`.
+
+Dense references are mandatory and existing latent/RGB quality limits remain
+unchanged. Reports include warmup time/memory and matched two-request diagnostic
+throughput comparisons. Require both `complete` and `passed` in `report.json`;
+timing is diagnostic, not an isolated performance result. See
+[Inference optimization](/docs/guides/inference-optimization) for compile,
+Graph, precision and batch compatibility.
+
 ## Benchmark modes
 
 - **`official-run`** — Execute the benchmark runtime declared by the manifest against prepared artifacts. Use only when the required evaluator, data, checkpoints, and services are ready.

diff --git a/docs/fumadocs/content/docs/reference/cli.zh.mdx b/docs/fumadocs/content/docs/reference/cli.zh.mdx
@@ -67,6 +67,12 @@ python -m worldfoundry zoo models --json
 
 **选哪条命令。** 日常可复现 / 模型页命令用 `worldfoundry-eval run <id> --pipeline.task-profile …`。只有 Workspace 视觉冒烟才用 `python -m worldfoundry.studio.workspace_job infer`。
 
+同一 Workspace 进程内，`WORLDFOUNDRY_STUDIO_BATCH_MAX_SIZE`（默认 `1`）和
+`WORLDFOUNDRY_STUDIO_BATCH_DELAY_MS`（默认 `5`）控制兼容原生 Wan T2V 与
+5B TI2V 常驻 worker 合批。准入要求并行作业槽、常驻权重及兼容的 Torch attention；
+编译批次使用 regional scope。配置在 worker 启动前生效，不跨独立 CLI 进程合批。
+精度与输入要求见[批处理准入、预热和取消契约](/zh/docs/guides/inference-optimization)。
+
 下面统一写 `worldfoundry-eval`，便于 shell history、job script 和问题报告保持一致。
 
 ## 每一层都有 Help
@@ -339,6 +345,14 @@ Python 客户端可通过 `worldfoundry.mcp.client.MCPClient` 和
 转换器接受 WAV/MP3 的 MIME 别名，其他音频格式会抛出 `ValueError`。
 文本与图像块保留原内容和 data URI；转换在本地完成，不调用 API provider。
 
+### 作业完成与取消
+
+持久化 command job 在 completion receipt 与恢复状态中保留最初的终止原因。
+超时保持 `failed` 及其 timeout error；用户取消保持 `cancelled`。
+
+POSIX 有界 benchmark command 在返回或向上传播中断前，清理其持有的进程组。
+Timeout payload 记录 `timed_out` 和 `kill_stuck`，后者表示清理尚未完成。
+
 Parser 才是 flag 的事实来源。请使用 `worldfoundry-eval <command> [<subcommand>] --help`，不要从无关模型或 benchmark recipe 复制参数。
 
 ## 常用工作流
@@ -477,6 +491,25 @@ worldfoundry-eval zoo benchmark-run \
 
 Data root、score directory、prompt manifest、judge credential、metric subset 等 benchmark 专属输入见对应的 [Benchmark Hub](/zh/docs/evaluation/benchmark-hub) 页面。
 
+### 6. 回放原生 Wan 执行路径与请求批处理
+
+在已准备本地 checkpoint 的 checkout 中运行：
+
+```bash
+python -m benchmarks.inference.wan_execution_preparation \
+  --checkpoint /path/to/Wan2.1-T2V-1.3B \
+  --output-dir tmp/wan-execution --request-batching
+```
+
+`--request-batching` 在单请求对照之外增加 A/B、B/A、A/B 批次。用 `--model-id`
+选择 `wan2.1-t2v-1.3b`、`wan2.2-t2v-a14b`、`wan2.2-ti2v-5b` 或
+`wan2.2-i2v-a14b`；图像条件回放还需 `--images IMAGE_A IMAGE_B`。
+
+Dense 参照必须执行，已有 latent/RGB 质量阈值保持不变。报告记录预热耗时/显存及
+同两个请求的诊断性吞吐对比。`report.json` 必须同时满足 `complete` 和 `passed`；
+计时仅作诊断，不是独占设备性能结果。Compile、Graph、精度与批处理兼容范围见
+[推理优化](/zh/docs/guides/inference-optimization)。
+
 ## Benchmark mode
 
 - **`official-run`** — 对准备好的 artifact 执行 manifest 声明的 benchmark runtime。只有 evaluator、数据、checkpoint 与服务都就绪后才使用。