docs(raw): publish 0.1.1 benchmark results

[bproject07]

Summary

• publish a second reproducible benchmark bound to immutable codex-raw-v0.1.1
• add 20 sanitized per-sample measurements plus exact release, tag, asset, binary, harness, environment, and method provenance
• update the benchmark overview and exact-tag reproduction commands

Run

• exact GitHub Windows x64 release asset; no local compilation
• gpt-5.4-mini, public prompt hello
• 2 warm-up turns and 20 measured turns (10 app-server + 10 api-server)
• sequential, no retries, no failures or timeouts

Privacy and interpretation

The v2 samples retain only order, mode, monotonic TTFT/total latency, input-token count, and output length. They contain no response text, output fingerprint, credential, account identifier, hostname, private address, or absolute local path. Unsalted output hashes were intentionally omitted because low-entropy outputs can be reconstructed from them.

The observed VM wall-clock interval is inconsistent with the summed monotonic samples by 242.5499 ms. This is recorded explicitly; no cause is asserted, and wall time is not used for latency calculations.

The two dated release runs are not presented as a controlled release-to-release performance comparison.

Verification

• independent recomputation of all medians, p90/p95, min/max, means, counts, ordering, and token invariants
• independent GitHub release/tag/asset/SHA256SUMS/binary/harness provenance audit
• staged JSON/hash binding and secret/privacy scan
• both PowerShell reproduction blocks parse and verify the exact tag commit and harness Git blob