test(rknn): add full-load qualification gates

Signed-off-by: samuel--hu <genius1@qq.com>

Author: samuel--hu

docs/en/guide/rk3576-rknn-development.md

@@ -123,7 +123,9 @@ python tools/rknn/prepare_validation_data.py \
   --output-dir yolov8-calibration --samples 32
 
 python tools/rknn/convert_model.py \
-  --spec config/rknn/models/yolov8.json --model yolov8-heads.onnx \
+  --spec config/rknn/models/yolov8.json \
+  --platform-profile config/rknn/platforms/rk3576.json \
+  --model yolov8-heads.onnx \
   --output yolov8-heads-int8.rknn --quantize \
   --dataset yolov8-calibration/dataset.txt
 ```

docs/en/tutorials/05-model-porting/model-porting.md

@@ -14,14 +14,14 @@ next: false
 | Who this is for | ML engineers and integration developers bringing a custom detector or classifier to CosmoEdge |
 | What you will accomplish | Evaluate runtime compatibility, convert and upload a model, configure parsing, and complete image, video, and sustained-run validation |
 | Prerequisites | Understand Pipelines and know the model input, output, preprocessing, postprocessing, and label order |
-| Estimated time | About 40–60 minutes for x86 ONNX; Sophon conversion commonly adds 30–60 minutes |
-| Device required | x86 requires an ONNX Runtime CosmoEdge build; Sophon requires a BM1688/CV186X device and matching conversion toolchain |
+| Estimated time | About 40–60 minutes for x86 ONNX; Sophon or Rockchip conversion commonly adds 30–60 minutes |
+| Device required | x86 requires an ONNX Runtime CosmoEdge build; Sophon and Rockchip require the actual target device and matching conversion toolchain |
 | Final acceptance result | The model loads, its output is parsed correctly, image and video results pass, and it runs without resource failure on the target device |
 
 Complete third-party integration in this order:
 
 1. Confirm support conditions and the model contract.
-2. Export ONNX; for Sophon, convert it again into a chip-specific `bmodel`.
+2. Export ONNX; convert it into a chip-specific `bmodel` for Sophon or `rknn` for Rockchip.
 3. Validate the artifact on the conversion host.
 4. Upload and configure the model.
 5. Run positive and negative image tests first.
@@ -39,12 +39,14 @@ output layout, and postprocessing are compatible with CosmoEdge.
 | --- | --- | --- | --- | --- |
 | x86 CPU | `.onnx` | `model.onnx` | ONNX Runtime CPU | x86_64 host and matching CosmoEdge build |
 | Sophon | `.bmodel` | `model.nn` | Sophon BMRT | BM1688 or CV186X; the artifact must target the actual chip |
+| Rockchip RKNN | `.rknn` | `model.rknn` | RKNN Runtime | RK3576 or RV1126B; the artifact must target the actual chip |
 
 `model.nn` is the internal file name in a CosmoEdge model package. It wraps the device model. When adding
 an individual Sophon model in the UI, select its `.bmodel`; do not rename an extension to `.nn`.
 
 PyTorch `.pt`, TensorFlow SavedModel, and other training-framework artifacts cannot be uploaded directly.
-Export them to ONNX first. Sophon deployments then convert ONNX into a chip-specific `.bmodel`.
+Export them to ONNX first. Sophon deployments then convert ONNX into a chip-specific `.bmodel`, while
+Rockchip deployments produce a chip-specific `.rknn`. RK3576 and RV1126B `.rknn` artifacts are not interchangeable.
 
 ### 1.2 Contracts Beyond the File Format
 
@@ -66,7 +68,7 @@ or runtime code.
 ### 1.3 Verified Capability vs Conditional Compatibility
 
 - **Directly supported by current code**: Add `.onnx` on x86, add `.bmodel` on Sophon, and import packages
-  containing `model.onnx` or `model.nn`.
+  containing `model.onnx` or `model.nn`; RKNN builds add `.rknn` and package it as `model.rknn`.
 - **Reference evidence in this repository**: a YOLOv8 detector has completed x86 ONNX import, live overlay,
   and event output.
 - **Still required on the target candidate**: validate your exact model, Sophon artifact, performance,
@@ -387,6 +389,32 @@ task acceptance, but it must not borrow another example's fixed shapes or hashes
 
 The Sophon Add Model page requires a `.bmodel` file.
 
+## Rockchip RKNN Path: Shared Backend, Target-Specific Artifacts
+
+RK3576 and RV1126B share one CosmoEdge RKNN inference implementation, Rockchip media interface, and model
+contract. Chip differences come from `config/rknn/platforms/<chip>.json`. A model spec does not hard-code a
+chip, but every conversion binds one platform profile, so the resulting `.rknn` remains target-specific and
+must not be copied between chips.
+
+The agent-assisted flow always enters through `scripts/agent/convert_model.sh` and `verify.sh`. It selects
+RKNN Toolkit2 from the task contract and freezes Python, wheel, platform profile, model spec, calibration
+set, and artifact hashes. RK3576 and RV1126B do not maintain separate conversion scripts. For manual
+diagnosis, follow the [RK3576 RKNN development guide](/en/guide/rk3576-rknn-development) and select the
+actual target profile:
+
+```bash
+python tools/rknn/convert_model.py \
+  --spec config/rknn/models/yolov8.json \
+  --platform-profile config/rknn/platforms/rv1126b.json \
+  --model yolov8-heads.onnx \
+  --output yolov8-rv1126b-int8.rknn \
+  --quantize --dataset calibration/dataset.txt
+```
+
+Conversion-host success is not device acceptance. Validate the matching Runtime/driver, numerical output,
+image and video postprocessing, OSD, rules, alerts, the 5 FPS target, and stability on the actual device,
+with results bound to the target chip and artifact SHA-256.
+
 ## 4. Upload and Configure the Model
 
 ### 4.1 Open Model Repository

docs/guide/rk3576-rknn-development.md

@@ -107,7 +107,9 @@ python tools/rknn/prepare_validation_data.py \
   --output-dir yolov8-calibration --samples 32
 
 python tools/rknn/convert_model.py \
-  --spec config/rknn/models/yolov8.json --model yolov8-heads.onnx \
+  --spec config/rknn/models/yolov8.json \
+  --platform-profile config/rknn/platforms/rk3576.json \
+  --model yolov8-heads.onnx \
   --output yolov8-heads-int8.rknn --quantize \
   --dataset yolov8-calibration/dataset.txt
 ```

docs/tutorials/05-model-porting/model-porting.md

@@ -14,14 +14,14 @@ next: false
 | 适合谁 | 需要把自有检测或分类模型接入 CosmoEdge 的算法工程师和集成开发者 |
 | 完成后能做什么 | 判断模型是否满足运行条件，转换并上传模型，配置解析参数，完成图片、视频和持续运行验证 |
 | 使用前提 | 已理解 Pipeline；掌握模型输入、输出、预处理、后处理和标签顺序 |
-| 预计时间 | x86 ONNX 路径约 40–60 分钟；Sophon 转换路径通常需要额外 30–60 分钟 |
-| 是否需要设备 | x86 路径需要 ONNX Runtime 版 CosmoEdge；Sophon 路径需要 BM1688/CV186X 设备及匹配转换工具链 |
+| 预计时间 | x86 ONNX 路径约 40–60 分钟；Sophon 或 Rockchip 转换路径通常需要额外 30–60 分钟 |
+| 是否需要设备 | x86 路径需要 ONNX Runtime 版 CosmoEdge；Sophon 与 Rockchip 路径需要目标芯片设备及匹配转换工具链 |
 | 最终验收结果 | 模型可加载、推理输出可解析、图片与视频结果正确，并在目标设备上持续运行无资源错误 |
 
 第三方模型接入按以下顺序完成：
 
 1. 确认支持条件和模型契约。
-2. 导出 ONNX；Sophon 设备再转换为目标芯片的 `bmodel`。
+2. 导出 ONNX；Sophon 设备转换为目标芯片的 `bmodel`，Rockchip 设备转换为目标芯片的 `rknn`。
 3. 在转换主机上检查产物。
 4. 上传并配置模型。
 5. 先做正负图片验证。
@@ -39,12 +39,14 @@ next: false
 | --- | --- | --- | --- | --- |
 | x86 CPU | `.onnx` | `model.onnx` | ONNX Runtime CPU | x86_64 主机和对应 CosmoEdge 构建 |
 | Sophon | `.bmodel` | `model.nn` | Sophon BMRT | BM1688 或 CV186X，转换产物必须匹配芯片 |
+| Rockchip RKNN | `.rknn` | `model.rknn` | RKNN Runtime | RK3576 或 RV1126B，转换产物必须匹配实际芯片 |
 
 `model.nn` 是 CosmoEdge 模型包中的内部文件名，封装的是设备侧模型；通过页面单独添加
 Sophon 模型时应选择 `.bmodel`，不要把文件扩展名手工改成 `.nn`。
 
 PyTorch `.pt`、TensorFlow SavedModel 或其他训练框架产物不能直接上传。它们必须先导出
-为 ONNX；Sophon 还要使用匹配工具链把 ONNX 转为目标芯片的 `.bmodel`。
+为 ONNX；Sophon 还要使用匹配工具链把 ONNX 转为目标芯片的 `.bmodel`，Rockchip 则转换
+为目标芯片的 `.rknn`。RK3576 与 RV1126B 的 `.rknn` 不是可互换模型。
 
 ### 1.2 格式之外还要匹配的契约
 
@@ -65,7 +67,7 @@ CosmoEdge 当前已有 `YOLOV8_DET` 等解析路径，但“任意 ONNX”并不
 ### 1.3 已验证能力与条件性兼容
 
 - **由当前代码直接支持**：x86 添加 `.onnx`、Sophon 添加 `.bmodel`，以及模型包中的
-  `model.onnx` / `model.nn`。
+  `model.onnx` / `model.nn`；RKNN 构建添加 `.rknn`，模型包中使用 `model.rknn`。
 - **仓库中已有参考证据**：YOLOv8 检测模型在 x86 ONNX 路径完成过模型导入、实时叠加
   和事件输出。
 - **仍需在目标候选版本上验证**：你的具体模型、Sophon 转换产物、性能、资源占用、
@@ -366,6 +368,29 @@ model_tool --info yolov8n_bm1688_f16.bmodel
 
 Sophon 添加模型页面会要求 `.bmodel` 文件，实际页面见下一节的添加模型表单。
 
+## Rockchip RKNN 路径：共享后端、目标专用产物
+
+RK3576 与 RV1126B 共用同一套 CosmoEdge RKNN 推理实现、Rockchip 媒体接口和模型契约；
+芯片差异由 `config/rknn/platforms/<chip>.json` 平台 profile 提供。模型 spec 不写死芯片，
+但每次转换必须绑定一个 profile，因此输出的 `.rknn` 仍是目标专用产物，不能跨芯片复制使用。
+
+智能体辅助流程统一使用 `scripts/agent/convert_model.sh` 与 `verify.sh`。执行器按任务合同选择
+RKNN Toolkit2，冻结 Python、wheel、平台 profile、模型 spec、校准集和产物哈希。RK3576 与
+RV1126B 不各自维护一套转换脚本。手工排障时可参考
+[RK3576 RKNN 开发说明](/guide/rk3576-rknn-development)，并把平台参数换成实际目标：
+
+```bash
+python tools/rknn/convert_model.py \
+  --spec config/rknn/models/yolov8.json \
+  --platform-profile config/rknn/platforms/rv1126b.json \
+  --model yolov8-heads.onnx \
+  --output yolov8-rv1126b-int8.rknn \
+  --quantize --dataset calibration/dataset.txt
+```
+
+转换主机通过不等于设备验收。必须在对应设备上继续验证 Runtime/驱动、数值输出、图片与视频
+后处理、OSD、规则、告警、5 FPS 目标以及稳定性；这些结果应绑定目标芯片和产物 SHA-256。
+
 ## 4. 上传并配置模型
 
 ### 4.1 进入模型仓库

tools/scenario-bench/README.md

@@ -76,11 +76,15 @@ node src/cli.js checkpoint \
   --max-gap-sec 120 \
   --min-fps-ratio 0.9 \
   --expected-preview-streams 4 \
-  --expected-decoder-backend rockchip-copy-out \
-  --expected-encoder-backend rockchip-copy-first
+  --expected-decoder-backend rockchip-mpp-rga \
+  --expected-encoder-backend rockchip-mpp-rga \
+  --min-rga-bound-native-int8-frames 1 \
+  --min-rknn-mpp-dmabuf-frames 1 \
+  --min-rknn-rga-crop-dmabuf-frames 1 \
+  --max-rga-bound-requantize-avg-ms 2
 ```
 
-输出结论只有三种：`IN_PROGRESS` 表示时长尚未达到且其余门禁正常（命令退出码 3），`FAIL` 表示连续性、负载、资源或原生媒体链路存在失败（退出码 1），`PASS` 仅在时长达到且全部必需检查通过时产生（退出码 0）。审计覆盖采样间断/新鲜度、固定路数、逐通道 FPS 与丢帧、CPU/内存/磁盘、内存池增长、MPP/RGA/RKNN 失败及计数器回退、延迟 Copy-out 记账、OSD/编码/发布帧流和 SRS 发布流；输入和候选身份文件的 SHA-256 会写入结果。
+输出结论只有三种：`IN_PROGRESS` 表示时长尚未达到且其余门禁正常（命令退出码 3），`FAIL` 表示连续性、负载、资源或原生媒体链路存在失败（退出码 1），`PASS` 仅在时长达到且全部必需检查通过时产生（命令退出码 0）。`gate-hours` 从首次观测到目标路数、活动路数和唯一通道数均达到配置负载时开始累计；登录、建连和部分路数爬坡只计入 `runAgeSec`，不计入验收使用的 `fullLoadCoverageSec`。审计覆盖采样间断/新鲜度、固定路数、逐通道 FPS 与丢帧、CPU/内存/磁盘、内存池增长、MPP/RGA/RKNN 失败及计数器回退、延迟 Copy-out 记账、OSD/编码/发布帧流和 SRS 发布流；针对原生 RKNN 路径还可要求 MPP DMA-BUF 输入、分类裁剪 DMA-BUF、全部 forward 的 INT8 直绑覆盖和必要符号位变换的平均耗时。输入和候选身份文件的 SHA-256 会写入结果。
 
 ## 创建场景包
 

tools/scenario-bench/src/cli.js

@@ -109,7 +109,7 @@ init-scenario options:
 checkpoint options:
   --input <file>       Running metrics.partial.json or completed metrics.json
   --output <file>      Atomic JSON checkpoint output path
-  --gate-hours <n>     Required continuous runtime, e.g. 12
+  --gate-hours <n>     Required observed full-load runtime; excludes login/ramp, e.g. 12
   --identity <file>    Optional immutable candidate identity key=value file
   --source-label <s>   Canonical remote/source path recorded in evidence
   --identity-label <s> Canonical identity path recorded in evidence
@@ -121,6 +121,10 @@ checkpoint options:
   --expected-decoder-backend <s> Required decoder backend identifier
   --expected-encoder-backend <s> Required encoder backend identifier
   --min-rga-bound-uint8-frames <n> Required fused RGA-to-RKNN UINT8 frames
+  --min-rga-bound-native-int8-frames <n> Required native INT8 bound frames with full RKNN accounting
+  --min-rknn-mpp-dmabuf-frames <n> Required RKNN frames sourced from MPP DMA-BUF
+  --min-rknn-rga-crop-dmabuf-frames <n> Required classifier crops sourced from DMA-BUF
+  --max-rga-bound-requantize-avg-ms <n> Maximum average native U8-to-INT8 transform latency
 `);
 }
 
@@ -751,6 +755,10 @@ async function runCheckpoint(args) {
     maxPoolGrowthBytes: optionalNumber(args, 'max-pool-growth-bytes'),
     poolGrowthWarmupSec: optionalNumber(args, 'pool-growth-warmup-sec'),
     minRgaBoundUint8Frames: optionalNumber(args, 'min-rga-bound-uint8-frames'),
+    minRgaBoundNativeInt8Frames: optionalNumber(args, 'min-rga-bound-native-int8-frames'),
+    minRknnMppDmaBufFrames: optionalNumber(args, 'min-rknn-mpp-dmabuf-frames'),
+    minRknnRgaCropDmaBufFrames: optionalNumber(args, 'min-rknn-rga-crop-dmabuf-frames'),
+    maxRgaBoundRequantizeAvgMs: optionalNumber(args, 'max-rga-bound-requantize-avg-ms'),
     expectedPreviewStreams: optionalNumber(args, 'expected-preview-streams'),
     expectedDecoderBackend: args['expected-decoder-backend'],
     expectedEncoderBackend: args['expected-encoder-backend'],