Add bounded exploratory API comparison and blind judging

[dKosarevsky]

Summary

• Add an explicitly exploratory 1,920-generation comparison across Sol, Terra,
  Luna and Astra, four arms, 24 frozen EN/RU cases and five repetitions.
• Add blinded Sol semantic judging, deterministic hard checks and per-model
  matched concise - if descriptive reports. Incomplete judgment coverage
  suppresses comparisons; failed/missing responses remain visible.
• Share one USD75 reservation/settlement ledger across generation and judging;
  retain raw attempts and partial artifacts, never retry authentication or
  uncertain calls. No resume or automatic rerun.
• Add a manual main-only first-attempt workflow, isolated credentials and locked
  dependencies. Existing smoke-pilot defaults and prior evidence are preserved.

This is not the governed confirmatory campaign or publication-ready evidence.
No measured improvement is claimed. The previous successful 32-response pilot
is not pooled into this run. No additional USD75 campaign is implicitly approved.

Verification

• Focused offline tests cover the new protocol/controller/workflow and pilot
  regression behavior, with synthetic responses and no provider credentials.
• Scoped Ruff checks and whitespace checks.
• Independent specification and quality reviews before integration.
• All protected-main CI jobs must pass before merge; no protection bypass.

Integration boundary

This PR is one exact-file commit on current main. It does not push the historical
experimental branch or activate the unfinished governed Runtime. A live API
dispatch is a separate, user-authorized manual action after merge.