docs(raw): publish 0.1.0 benchmark provenance

[bproject07]

Summary

• explain the core concept near the top of the README: Codex Raw removes the Codex agent layer, not account controls
• publish the first release-bound benchmark with sanitized per-sample data and exact provenance
• include BENCHMARKS.md in future platform archives
• fix future draft verification by resolving the numeric GitHub release ID before REST asset checks

Benchmark identity

• immutable release: codex-raw-v0.1.0
• tag commit: 99cc140866c4f66da6d8f283a013e6b43e4458e9
• Windows archive SHA-256: c5551071b41245c33d6406a2a2f20c95fc8ff91bcd7468ab41b701a30c661d08
• extracted binary SHA-256: 7a5d8b8a1b7144a2a6efd619b53a477a251acc19c14a6e06f85e8e23baa15208
• exact tagged harness Git blob: 2c92994556a50dd2c83be9c844ad41d141a6fbc8
• one uninterrupted run: 2 warm-ups plus 20 measured turns, no retries, failures, or timeouts

The report explicitly treats these numbers as a dated observation, records the API tail sample and output difference, and makes no general speed claim.

The immutable 0.1.0 release and tag are not changed by this PR.

Validation

• Prettier 3.5.3: pass
• Actionlint 1.7.12: pass
• YAML parse with duplicate-key rejection: pass
• Bash syntax for all workflow Bash blocks: pass
• Markdown relative-link check: pass
• JSON parse, counts, ordering, input-token invariant, and digest linkage: pass
• independent recomputation of median/P90/P95/min/max: pass
• Gitleaks 8.30.1 staged scan: 0 findings