bench: record initial native baseline

Author: myfreess

README.mbt.md

@@ -92,14 +92,16 @@ moon info
 ```
 
 The ASan entry point is `scripts/run-asan`. The performance gate is
-`scripts/bench-gate`; it requires a clean committed baseline so its manifest
-can contain a real commit SHA. See [`benchmarks/README.md`](benchmarks/README.md).
+`scripts/bench-gate`. Its committed
+[`benchmarks/baseline.json`](benchmarks/baseline.json) binds the fixed
+workloads and 20 raw samples to clean commit
+`29f4ff26dcb507715a0bfdacd58a0a3b48f3c776`. See
+[`benchmarks/README.md`](benchmarks/README.md).
 
 The generated per-item upstream inventory is
 [`docs/upstream-coverage.json`](docs/upstream-coverage.json). Public interfaces
 are reviewed through every `pkg.generated.mbti`; DTLS record ciphers, PRF/key
 schedule helpers, crypto providers, parsers, and replay state remain internal.
 
-The project is dual-licensed under MIT or Apache-2.0. It remains pre-1.0 until
-the first clean release commit can provide the benchmark baseline's truthful
-commit identity.
+The project is dual-licensed under MIT or Apache-2.0. It remains pre-1.0 while
+the reviewed public API is intentionally experimental.

UPSTREAM.md

@@ -111,8 +111,6 @@ The repository contains:
   stubs while disabling the bundled allocator;
 - seven native release benchmarks and a self-testing interleaved regression
   gate;
+- a five-warmup, twenty-sample raw benchmark baseline tied to clean commit
+  `29f4ff26dcb507715a0bfdacd58a0a3b48f3c776`;
 - generated `.mbti` files for public-interface review.
-
-`benchmarks/baseline.json` must be recorded from the first clean release
-commit; this repository currently has no commit from which a truthful
-baseline SHA can be recorded.

benchmarks/README.md

@@ -4,6 +4,11 @@
 The gate checks the benchmark source hash, so workload changes require a
 reviewed baseline update.
 
+The current manifest was recorded from clean commit
+`29f4ff26dcb507715a0bfdacd58a0a3b48f3c776` with five warmups and twenty
+samples. It retains the runner fingerprint (machine, CPU model, and optional
+runner ID), aggregates, and complete raw benchmark output.
+
 Record a baseline only from a clean commit:
 
 ```sh
@@ -27,5 +32,9 @@ fixed-workload throughput falls by more than 20%. Set
 `RTC_BENCH_RUNNER_ID` to a stable dedicated-runner identifier before both
 recording and comparing.
 
+Both worktrees use the MoonBit toolchain installed in the current environment.
+The manifest intentionally does not record or require a specific MoonBit
+version.
+
 Use `scripts/bench-gate self-test` to validate parsing and threshold logic
 without running benchmarks.