Verify the harness

Read AGENTS.md, INVARIANTS.md, current CI and nearest tests. Identify the real checkout, source SHA, dirty state, binary and architecture. Preserve operator homes/services. Inspect fixtures and current test assertions.

Choose checks by the changed contract:

• Backend: run the changed command, route or function and inspect its effects; use the scoped local checks in AGENTS.md. Broad regression tests belong in CI.
• Contracts: use cgagentharness-invariant-guard for current test mapping, including ownership, memory, MCP, web, reload, schedules and notifications. Documentation edits select invariant_guard::markdown_stays_on_the_docs_budget.
• Native execution: tests/macos_cargo.rs and process lifecycle tests need real macOS Seatbelt; sandbox permission errors are not application regressions.
• Desktop: use cgagentharness-release for packaging order and provenance; select desktop checks and scripts/test-desktop-backend.py cases that exercise the changed boundary against the packaged backend.
• Model/runtime: inspect actual inventory and exact configured model tag; do not equate similarly named GGUF/MLX variants or turn on cloud fallback for a smoke test.

For live acceptance, use an owned temporary home and unique port. Record the known binary, environment and listener, issue harmless local chat, then stop it. Fake model/provider tests prove protocol behavior, not availability or live quality. App launch alone is not UI acceptance. Record only native UI interactions actually observed, in the PR body.

Keep a failing-case reproducer for corrections; diagnose baseline failures before attributing them to the candidate. Report commands, exit statuses, tested SHA/architecture, artifact hashes and remaining limits. A zero-test selection is NOT RUN. Follow AGENTS.md for preserving assertions and operator data/services.