knowledge: merge main into the 2026-09-10 batch and curate it

The batch was distilled on the SDK pinned before #142 and against a main
that predates #140/#145, so it reused two IDs those batches landed and its
audit ran against the retired baseline. Merge main in and curate:

- VOMNI-CFG-1s (CLI-only negative alias) -> VOMNI-CFG-1t and
  MCPMO-2e (audio_embeds registration) -> MCPMO-2f: both IDs already head
  other rules on main; every ID now heads exactly one section tree-wide.
- The configuration entry page crossed the 32 KiB split cap once main's
  VOMNI-CFG-1s and this batch's 1t both sat on it. Both engine-args rules
  move to the existing topic page rules-legacy-engine-args.md (the owner's
  engine-args topic), the entry page's Direct table links there, and the
  nearest _index.md row is extended. No inbound anchors referenced them.
- Every cited PR added to its page's `sources:` frontmatter; Direct table
  rows for OMNI-CI-2g/2h, DIFF-1af, MCPMO-2f, VOMNI-CFG-1t and _index.md
  rows for benchmark, ci, configuration and diffusion.

The review gate also noted that the SDK deduplicates heading IDs within a
single section, so a section repeating its own ID in a nested heading is
not rejected; that is an SDK change and ships separately.

Validators: check_knowledge_tree 0 errors, check_wiki_lint 0 errors,
test_thin_mcp_server and test_knowledge_source green.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018MAFmwP3u5qAMDgHfM6LFi

Author: tzhouam

adapters/vllm_omni/release_baseline.yaml

@@ -2,7 +2,7 @@ schema_version: 1
 
 upstream:
   repository: vllm-project/vllm-omni
-  previous_audited_sha: 1015b1ced073195ed19f0610264dfdf8a8150e6c
+  previous_audited_sha: 66bbebd9d3ef669b93fd7e39587672b20efc5933
   audited_ref: main
   audited_sha: 44d3ae100afa8411770b8b4d9442318bc5e897b3
 

doc/architecture/SPEC/sdk.md

@@ -1,6 +1,7 @@
 # sdk/ —— 规范
 
-<!-- verified-against: 2026-09-08 -->
+<!-- verified-against: 2026-09-10 -->
+
 
 `Python SDK v1 · 跨仓库唯一 typed 边界 · refactor-status: ok`
 
@@ -56,12 +57,17 @@
   `free_bytes`/`free_lines`，与 `check_knowledge_tree.py` 的拆分门对齐）；prompt 把
   事件放进唯一 `<untrusted_data>` fence（`KnowledgeEvidenceEvent.diff_excerpt` 是宿主
   选取的 diff 片段，SDK 以每事件 8 KiB、每批 160 KiB 截断，先到先得），proposal shape、rule ID、heading、source
-  citation、目标页、重复 ID 与页面剩余容量都由 SDK 机械校验——超出目标页剩余容量的
+  citation、目标页、重复 ID 与页面剩余容量都由 SDK 机械校验——rule ID 全树唯一：
+  任何 catalog 页（所有仓库与 general）、任何 heading 层级已使用的 ID，以及 section
+  内嵌套 heading 携带的 ID，都被拒绝，同批两个 proposal 也不得共用 ID；超出目标页剩余容量的
+
   proposal 以 `page full` 在 validator 之前被拒绝，宿主据此改路由而不是整批回滚。
   `proposal_id` 同时绑定 batch、输入下标、repository、section、sources 与目标页
   SHA，宿主不能在 model call 后静默改写已接纳 proposal。
 - **知识 apply 是 append-only transaction**：只追加完整 rule section 并更新唯一
-  `updated:` frontmatter；写前复核 page SHA。固定且按序执行
+  `updated:` frontmatter；写前复核 page SHA，并在锁内再次全树复核 rule ID 唯一性
+  （并行校验的两批不能把同一新 ID 落到不同页）。固定且按序执行
+
   `knowledge/tools/check_knowledge_tree.py`、`knowledge/tools/check_wiki_lint.py`；
   validator 缺失则写前 fail closed，执行失败/超时则逐 byte rollback 全部目标页。
   同一 work checkout 的 writer 以 process 内 mutex 与位于系统临时目录的

knowledge/repos/vllm-omni/benchmark/_index.md

@@ -21,7 +21,7 @@ sources: [benchmarks/tts/bench_tts.py, benchmarks/tts/model_configs.yaml, "PR #6
 
 | 遇到什么 | 查看哪里 |
 |---|---|
-| 新增/审查 perf JSON、warmup 覆盖、realtime artifact 或 Hub dataset fallback | [benchmark rules](rules.md) |
+| 新增/审查 perf JSON、warmup 覆盖、realtime artifact、Hub dataset fallback、per-request timeout 或 `/v1/videos` 轮询预算 | [benchmark rules](rules.md) |
 | 查看旧 benchmark 总览 | [overview](overview.md) |
 | 运行 AR graph、Hunyuan 或 MiniCPM-o Omni-DuplexEval benchmark | [benchmark guides](guides/_index.md) |
 | 调查 profiling、模型加载和性能验证错误 | [benchmark incidents](incidents/_index.md) |

knowledge/repos/vllm-omni/benchmark/rules.md

@@ -4,7 +4,7 @@ created: 2026-09-05
 updated: 2026-09-10
 type: rule
 tags: [vllm-omni, benchmark]
-sources: ["PR #6817", tests/dfx/perf/scripts/run_benchmark.py, tests/benchmarks/test_omniinteract.py]
+sources: ["PR #6817", tests/dfx/perf/scripts/run_benchmark.py, tests/benchmarks/test_omniinteract.py, "PR #7130", "PR #7259"]
 confidence: high
 ---
 

knowledge/repos/vllm-omni/ci/_index.md

@@ -15,7 +15,8 @@ sources: [.buildkite/cuda/pipeline.yml, docs/contributing/ci/test_system_overvie
 - 处理 `@hardware_test`/`hardware_marks`、SKU 与 `cards_N` 自动 marker 或按卡数切分时，先看
   [测试分级与 markers](guides/test-tiers.md)，再按 OMNI-CI-1a 审核 lane 的真实收集与执行。
 - Qwen3-TTS Base/CustomVoice 的 ready/merge oracle 与 source dependency 见
-  [TTS CI 规则](rules-tts.md)。
+  [TTS CI 规则](rules-tts.md)（含 task↔checkpoint 配对与 speaker-embedding miss-EOS retry）。
+
 
 ## MiniCPM-o 4.5 full-payload regression coverage
 
@@ -209,10 +210,10 @@ sources: [.buildkite/cuda/pipeline.yml, docs/contributing/ci/test_system_overvie
 | 遇到什么 | 查看哪里 |
 |---|---|
 | 审查硬件 lane（含 MiniMax-H3 DLO DP2 ready smoke 的证据边界）、回归 fence、CI 工具供应链、ASR 文本比较或 xdist/shared fixture | [CI rules](rules.md) |
-| AMD/ROCm timeout、quarantine、memory signal、Qwen3-TTS argv 或 Qwen3-Omni control-plane fixture | [AMD/ROCm CI rules](rules-amd.md) |
+| AMD/ROCm timeout、quarantine、memory signal、single-card diffusion job 的 multi-card marker 排除、Qwen3-TTS argv 或 Qwen3-Omni control-plane fixture | [AMD/ROCm CI rules](rules-amd.md) |
 | CUDA L4 Kubernetes preset、GPU-count shard、resource/retry policy | [L4 Kubernetes CI rules](rules-l4-k8s.md) |
 | Whisper 转写 helper 的 GPU 首选、16 GiB 门槛、CPU fallback 或 CUDA ready/merge source dependency | [Whisper 转写 CI 规则](rules-whisper-transcription.md) |
 | 查看仓库特有 CI 陷阱 | [CI guides](guides/_index.md) |
 | 调查历史 CI 失败 | [CI incidents](incidents/_index.md) |
 
-- [CI 并行测试与 engine fixture 合同](rules-test-fixtures.md)：`OMNI-CI-2b`, `OMNI-CI-2c`, `OMNI-CI-2d`。
+- [CI 并行测试与 engine fixture 合同](rules-test-fixtures.md)：`OMNI-CI-2b`, `OMNI-CI-2c`, `OMNI-CI-2d`, `OMNI-CI-2h`。

knowledge/repos/vllm-omni/ci/rules-amd.md

@@ -4,7 +4,7 @@ created: 2026-09-05
 updated: 2026-09-10
 type: rule
 tags: [vllm-omni, ci]
-sources: ["PR #6704", "PR #6830", "PR #6884", .buildkite/amd/, tests/helpers/clean.py, tests/helpers/stage_config.py, tests/buildkite/test_amd_pipeline.py, tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py]
+sources: ["PR #6704", "PR #6830", "PR #6884", .buildkite/amd/, tests/helpers/clean.py, tests/helpers/stage_config.py, tests/buildkite/test_amd_pipeline.py, tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py, "PR #7234"]
 confidence: high
 ---
 

knowledge/repos/vllm-omni/ci/rules-test-fixtures.md

@@ -4,7 +4,7 @@ created: 2026-08-23
 updated: 2026-09-10
 type: rule
 tags: [vllm-omni, ci]
-sources: ["PR #3422", "PR #5074", "PR #5255", "PR #5310", "PR #5402", "PR #5524", "PR #5543", "PR #5670", "PR #5713", "PR #5780", "PR #5823", "PR #5836", "PR #5957", "PR #5976", docker/Dockerfile.ci, docker/Dockerfile.xpu, .buildkite/intel/scripts/run-xpu-test.sh, .buildkite/cuda/test-merge.yml, .buildkite/cuda/test-ready.yml, "PR #5845", "PR #5872", "PR #6008", "PR #6048", "PR #6056", "PR #6096", "PR #6102", "PR #6202", "PR #6208", "PR #6273", "PR #6293", "PR #6311", "PR #6339", "PR #6343", "PR #6468", "PR #6523", "PR #6613", "PR #6555", "PR #6650", .buildkite/common/scripts/run_cov_split.sh, pyproject.toml, tests/helpers/tests/test_mark.py, tools/pre_commit/check_test_marks.py, .buildkite/common/scripts/upload_pipeline.py, .buildkite/cuda/test-nightly.yml, .buildkite/cuda/test-weekly.yml, .buildkite/npu/test-npu-nightly.yml, .pre-commit-config.yaml, tests/helpers/clean.py, tests/helpers/client.py, tests/helpers/mark.py, tests/helpers/runtime.py, tests/helpers/stage_config.py, tests/buildkite/test_upload_pipeline.py, tests/dfx/perf/scripts/run_benchmark.py, tests/dfx/perf/tests/test_minicpmo_4_5.json, tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json, tests/dfx/perf/tests/test_qwen_image_vllm_omni.json, tests/dfx/stability/, tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py, tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py, tests/e2e/online_serving/test_flux_kontext_expansion.py, tests/e2e/online_serving/test_minicpmo_4_5.py, tests/e2e/online_serving/test_minicpmo_4_5_duplex.py, tests/e2e/online_serving/test_minicpmo_4_5_expansion.py, tests/e2e/online_serving/test_qwen_image_expansion.py, tests/e2e/online_serving/test_minimax_h3_dlo_dp2_t2va.py, tests/model_tests/diffusion/diff_model_builders.py, tests/model_tests/diffusion/model_settings.py, tests/model_tests/diffusion/test_alignment.py, tools/nightly/run_nightly_jobs.sh, tools/pre_commit/check_tts_adapter.py, tests/tools/test_check_tts_adapter.py, .buildkite/amd/scripts/bootstrap-amd-omni.sh, .buildkite/amd/test-amd-merge.yml, .buildkite/amd/test-amd-ready.yml, tests/diffusion/distributed/test_tensor_parallel.py, tests/diffusion/offloader/test_diffusion_layerwise_offload.py, "PR #6704", tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json, tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json, "PR #6743", "PR #6696", vllm_omni/benchmarks/metrics/metrics.py, vllm_omni/benchmarks/patch/patch.py, tests/benchmarks/metrics/test_metrics.py, tests/benchmarks/patch/test_patch.py, "PR #6674", docker/Dockerfile.npu, docker/Dockerfile.npu.a3, docker/Dockerfile.npu.ci, docker/Dockerfile.npu.ci.a3, "PR #6818", "PR #6830", "PR #6884", tests/diffusion/conftest.py, tests/diffusion/attention/test_flash_attn.py, tests/buildkite/test_amd_pipeline.py, tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py, "PR #6947"]
+sources: ["PR #3422", "PR #5074", "PR #5255", "PR #5310", "PR #5402", "PR #5524", "PR #5543", "PR #5670", "PR #5713", "PR #5780", "PR #5823", "PR #5836", "PR #5957", "PR #5976", docker/Dockerfile.ci, docker/Dockerfile.xpu, .buildkite/intel/scripts/run-xpu-test.sh, .buildkite/cuda/test-merge.yml, .buildkite/cuda/test-ready.yml, "PR #5845", "PR #5872", "PR #6008", "PR #6048", "PR #6056", "PR #6096", "PR #6102", "PR #6202", "PR #6208", "PR #6273", "PR #6293", "PR #6311", "PR #6339", "PR #6343", "PR #6468", "PR #6523", "PR #6613", "PR #6555", "PR #6650", .buildkite/common/scripts/run_cov_split.sh, pyproject.toml, tests/helpers/tests/test_mark.py, tools/pre_commit/check_test_marks.py, .buildkite/common/scripts/upload_pipeline.py, .buildkite/cuda/test-nightly.yml, .buildkite/cuda/test-weekly.yml, .buildkite/npu/test-npu-nightly.yml, .pre-commit-config.yaml, tests/helpers/clean.py, tests/helpers/client.py, tests/helpers/mark.py, tests/helpers/runtime.py, tests/helpers/stage_config.py, tests/buildkite/test_upload_pipeline.py, tests/dfx/perf/scripts/run_benchmark.py, tests/dfx/perf/tests/test_minicpmo_4_5.json, tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json, tests/dfx/perf/tests/test_qwen_image_vllm_omni.json, tests/dfx/stability/, tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py, tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py, tests/e2e/online_serving/test_flux_kontext_expansion.py, tests/e2e/online_serving/test_minicpmo_4_5.py, tests/e2e/online_serving/test_minicpmo_4_5_duplex.py, tests/e2e/online_serving/test_minicpmo_4_5_expansion.py, tests/e2e/online_serving/test_qwen_image_expansion.py, tests/e2e/online_serving/test_minimax_h3_dlo_dp2_t2va.py, tests/model_tests/diffusion/diff_model_builders.py, tests/model_tests/diffusion/model_settings.py, tests/model_tests/diffusion/test_alignment.py, tools/nightly/run_nightly_jobs.sh, tools/pre_commit/check_tts_adapter.py, tests/tools/test_check_tts_adapter.py, .buildkite/amd/scripts/bootstrap-amd-omni.sh, .buildkite/amd/test-amd-merge.yml, .buildkite/amd/test-amd-ready.yml, tests/diffusion/distributed/test_tensor_parallel.py, tests/diffusion/offloader/test_diffusion_layerwise_offload.py, "PR #6704", tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json, tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json, "PR #6743", "PR #6696", vllm_omni/benchmarks/metrics/metrics.py, vllm_omni/benchmarks/patch/patch.py, tests/benchmarks/metrics/test_metrics.py, tests/benchmarks/patch/test_patch.py, "PR #6674", docker/Dockerfile.npu, docker/Dockerfile.npu.a3, docker/Dockerfile.npu.ci, docker/Dockerfile.npu.ci.a3, "PR #6818", "PR #6830", "PR #6884", tests/diffusion/conftest.py, tests/diffusion/attention/test_flash_attn.py, tests/buildkite/test_amd_pipeline.py, tests/e2e/offline_inference/test_qwen3_omni_colocate_async.py, "PR #6947", "PR #7225"]
 confidence: high
 ---
 

knowledge/repos/vllm-omni/ci/rules-tts.md

@@ -1,10 +1,10 @@
 ---
 title: "vLLM-Omni TTS CI 规则"
 created: 2026-09-05
-updated: 2026-09-05
+updated: 2026-09-08
 type: rule
 tags: [vllm-omni, ci]
-sources: ["PR #6861", .buildkite/cuda/test-merge.yml, .buildkite/cuda/test-ready.yml, .buildkite/amd/test-amd-merge.yml, tests/e2e/online_serving/test_qwen3_tts_base.py]
+sources: ["PR #6861", .buildkite/cuda/test-merge.yml, .buildkite/cuda/test-ready.yml, .buildkite/amd/test-amd-merge.yml, tests/e2e/online_serving/test_qwen3_tts_base.py, "PR #7054", "PR #6924"]
 confidence: high
 ---
 
@@ -19,3 +19,18 @@ confidence: high
 - 禁止：用 Ready dummy-weight 绿灯证明 Base real-weight/EOS 行为，或把 pending merge job 写成已证明通过。
 - 验收：Ready `core_model` collect-only 对 Base 为空；CUDA/AMD merge marker matrix 仍含 Base offline+online；
   source-file mutation 分别验证上述 CUDA scope。^[PR #6861] ^[issue #6855]
+
+## OMNI-CI-3g — Qwen3-TTS perf/CI case 必须把 task 与 checkpoint 变体成对绑定
+
+- 触发：修改 Qwen3-TTS benchmark registry、DFX perf JSON、nightly/merge TTS workload，或新增 `voice_clone` / `default_voice` / `voice_design` case。
+- 强制：每个 task 必须绑定与之匹配的 checkpoint 变体并在 server params 中显式启动该变体：`Base` 只服务 `voice_clone`，`CustomVoice` 只服务 `default_voice`，`VoiceDesign` 只服务 `voice_design`。切换 task 时必须重启到匹配 checkpoint；仅改 benchmark request body 或 `--model` 字段而不换 server，不构成有效 coverage。
+- 强制：registry/README/CI JSON 三处必须同步声明支持矩阵，防止 validation 已收紧后仍把旧 task 发到错误 checkpoint，令 suite 以 HTTP 400 失败。
+- 禁止：沿用“相邻变体应该也能跑”的历史假设；把 one checkpoint covers all 三种 task 写进 perf smoke；或只改文档不改实际 perf case。
+- 验收：至少一条实际 perf/CI 配置分别覆盖 Base、CustomVoice、VoiceDesign 的目标 task，并证明错误配对会在 serving validation 提前失败；通过结果只绑定所配对的那个 checkpoint-task workload。^[PR #7054]
+
+## OMNI-CI-3f — Qwen3-TTS Base speaker-embedding 测试不得钉死 `max_new_tokens`
+
+- 触发：修改 `test_qwen3_tts_speaker_embedding`、Base dummy/x-vector embedding E2E，或依赖 serving miss-EOS retry 的 TTS CI。
+- 强制：非流式请求省略 `max_new_tokens`，以便 serving 在 miss-EOS 时丢弃不完整生成并按 [Q3TTS-4b](../models/qwen3-tts/rules.md) 重试一次。raw PCM streaming 无法在 serving 内重试时，测试层可对 codec-EOS HTTP 500 或 `RemoteProtocolError` 整请求最多再试一次；无关 500 不得重试。
+- 禁止：为“加快测试”固定 `max_new_tokens` 从而关掉 serving retry；把测试层成功重试写成 miss-EOS 已不是 failure；或对任意 500 盲目重试。
+- 验收：CPU 单测覆盖 EOS-miss 重试、无关 500 不重试、`extra_retries=0` 上抛；GPU real-weight 路径在省略 max 的非流式与带一次 client retry 的 streaming 上通过。^[PR #6924]