feat: add bounded exploratory API pilot

[dKosarevsky]

Summary

• Add a standalone, standard-library exploratory API pilot: four models (including GPT-6 Astra), four arms, one EN/RU scenario, one repetition, at most 32 generation requests and no retries.
• Reserve exposure before each call under a dated USD 5 local budget; require complete usage/cache/reasoning accounting before settlement. Preserve raw responses and partial failure evidence; stop on uncertainty.
• Add a pinned, manually dispatched, main-only, first-attempt GitHub workflow. The benchmark-live secret is scoped to the live step; offline preflight is credential-free and artifacts are retained on failure.
• This PR contains only five pilot files on current main. It does not ship the experimental campaign branch, change GitHub protection settings, or claim measured efficiency gains.

Validation

• Focused offline launcher and workflow regression tests, including failure paths and encoded credential echoes.
• Direct standard-library unittest entry points, Ruff lint/format, workflow YAML parsing and offline CLI dry run.
• Independent specification and code-quality review.
• Full repository CI remains required before merge. No paid provider requests have been made during implementation or tests.

Operational boundary

The first live dispatch is a separately announced exploratory pilot, not the quality-gated public campaign. The local exposure cap assumes the documented conservative input bound and the 2026-09-18 standard global price snapshot; it is not a provider invoice guarantee or an account-wide spend cap. No semantic judge or comparative claim is included.