test(benchmarks): automate Promptfoo setup

[tulayha]

Related issue

None.

Summary

• add a single npm --prefix benchmarks/codex-mcp run setup entry point for the Codex MCP ablation
• automate dependency installation, isolated Codex authentication, pinned LongVALE download and verification, model preparation, indexing, saved environment configuration, and preflight
• install only the host-specific Promptfoo libsql binding while continuing to omit unrelated optional providers
• make repeat setup runs reuse prepared state and skip an already complete pilot index
• replace the manual PowerShell setup sequence in the benchmark guide with the automated workflow

This is internal-only benchmark tooling. It does not change a public product interface, storage format, or model format, and existing repositories do not need to be rebuilt.

Validation

• npm --prefix benchmarks/codex-mcp run test:setup — 5 Node unit tests passed for version checks, cross-platform paths, environment serialization, index completeness, and native binding selection
• npm --prefix benchmarks/codex-mcp run check — Promptfoo 0.122.2 loaded its Windows libsql binding and reported Configuration is valid.
• npm --prefix desktop run check — TypeScript, ESLint, 47 Vitest tests, and the Vite production build completed under Node 24.20.0
• npx --yes markdownlint-cli2@0.23.2 — completed with no findings
• git diff --check — completed with no whitespace errors

The full setup was not run because it requires the 1,063,510,782-byte LongVALE archive, model preparation, and a dedicated interactive Codex login. No Codex inference was run. Lychee was not run locally because Docker Desktop was unavailable; the repository documentation workflow remains responsible for that check on the PR.