[zhangzhenghao]

Summary

• Add clap bench for serial text, audio embeddings and waveform-frame inference using pinned Google Benchmark v1.9.4.
• Forward native benchmark options after --, including filtering, repetitions, listing and JSON/CSV output; fail on inference/report errors and unmatched filters.
• Measure full synchronous public C API calls (preprocessing and output allocation/free included), excluding weight loading, audio decoding and one warmup; report wall-clock milliseconds and item throughput.
• Keep Google Benchmark static/PIC and private to the CLI; core-only builds and exported core SDK dependencies remain unaffected.
• Add generated tiny-model CLI and packaging regressions (README unchanged).

Verification

All local builds and compilation temporary files are on /dev/sdb1 under /mnt/openclaw/clap-bench-dev.

• Final static full CTest suite: 37/37 passed; new benchmark CLI and packaging tests separately re-run after test-portability fixes.
• Shared build and final full CTest suite: 36/36 passed.
• Core-only configuration, build and packaging/install checks verified locally.
• Installed shared CLI exercised with real generated-model inference.
• No published-checkpoint performance or GPU measurements claimed; numerical/inference checks use deterministic generated fixtures.

Remote CI

All 9 checks passed on head e222bf04eef7ac4a9a41ac80a34f604ffba05ca7 (run 42550921790): formatting; native build/test/install and applicable Python/Node tests on Linux x64/ARM64, macOS ARM64, Windows x64/ARM64; Rust checks on Linux, macOS and Windows.

Earlier CI findings were fixed in the new test infrastructure: honor custom executable install directories and avoid minimum-duration assertions against native benchmark calibration. The transient Hugging Face download failure in the unchanged Node suite did not recur on the latest head. README remains unchanged.