refactor(optimizer): delegate model calls to runtime owners

[drewstone]

Closes #537

What changed

• replace Eval-owned provider credentials and HTTP calls with one caller-owned canonical model callback
• translate official GEPA, SkillOpt, and DSPy OpenAI-compatible child requests at the Eval boundary
• require a typed canonical response, exact cost receipt, and finite execution evidence for every invoked call
• persist callback identity and model execution observations in benchmark artifacts
• remove the legacy direct-provider path and public benchmark credential flags

Contract

• requests carry a stable call ID, deeply immutable canonical ChatRequest, abort signal, and endpoint format
• responses carry canonical ChatResponse, CostReceiptInput, and JSON execution evidence
• input tokens exclude cached tokens; cache-write tokens remain separate
• owner failures are returned through the local child endpoint without leaking provider bodies or credentials

Proof

• pnpm test: 4,742 passed, 3 skipped
• pnpm typecheck; pnpm typecheck:examples; pnpm lint
• pnpm build; pnpm verify:package
• full pinned GEPA + SkillOpt Python suite passed
• official GEPA and SkillOpt request fixtures captured and replayed, including SkillOpt dynamic system-role messages
• DSPy bridge: 35 passed
• local fake-provider boundary suite: 43 passed

[tangletools]

✅ Auto-approved drewstone PR — f66212b7

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-04T02:05:19Z}

[tangletools]

✅ Auto-approved drewstone PR — 03c46190

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-08-04T02:08:41Z}