Add offline fixtures and side-effect-safe probe infrastructure #9

Problem

Several probes rely on live arXiv access and actual GitHub issue creation. Network availability, account state, and changing external data contaminate model scores; live issue creation also creates unwanted side effects.

Goal

Run all benchmark probes against deterministic local substitutes with explicit setup, reset, and cleanup behavior.

Scope

• Provide fixed arXiv-like search/detail fixtures that exercise the research skill interface.
• Provide a disposable local GitHub-compatible issue target or a narrow mock that records created issues.
• Add deterministic fixture data for each probe, including expected facts to propagate through workflow chains.
• Add lifecycle commands for setup, reset, and cleanup.
• Document the supported isolation boundary and prohibit calls to production external services during a benchmark run.

Acceptance Criteria

• The full suite runs with outbound network access disabled after images/dependencies are available.
• No probe can create or modify a real GitHub issue or repository.
• Fixture responses are versioned and have stable IDs/content across runs.
• Workflow-chain probes can validate citations and conclusions against fixture facts.
• Setup/reset/cleanup leave no state that affects a subsequent run.
• A test verifies that external service endpoints are not contacted.

Non-Goals

• Perfect emulation of arXiv or GitHub.
• Evaluating browsing quality against the live web.

Implementation Notes

Favor lightweight stdlib services or static fixtures over new runtime dependencies. Fixture interfaces should match the tools exposed to Hermes closely enough to preserve the intended skill behavior.

Dependencies

• Depends on Restore a self-contained, runnable benchmark pipeline #5 for the stable runner and run lifecycle.
• Enables deterministic scoring work in Make HARP scoring deterministic and behaviorally valid #6.