[swerner]

Default native agent stages could accept a model response after their timeout: dispatch selected native API, but lowering marked the stage as handler-managed, so neither the session nor the driver enforced the deadline. Treat an omitted agent backend as executor-enforced, using the same interview-aware timer as explicit backend="api".

Add regressions for backend resolution (default, node, graph, and stylesheet), a three-second fake response against a one-second budget, and deadline cancellation of a child agent's tool shell and descendant process. Native session tests now exercise the default backend. ACP and prompt-stage policies are unchanged.

Fixes #101.

Validation:

• Confirmed the new policy and delayed-response regressions fail before the fix.
• Focused frontend, native-agent, ACP timeout, descendant cleanup, interview-budget, and retry tests: 100 passed.
• Full workspace suite: 1,750 passed, 9 skipped; documentation examples passed. Nextest reported a non-failing leak warning in the existing petri-runtime::host_actions::failed_action_cleanup_preserves_the_host_workspace_until_retry test, which does not exercise agent lowering.
• Formatting, Clippy, whitespace, vendored-bundle verification, workflow audit, and remote Docker artifact check passed.
• Release build passed; the complete mise run check gate exited successfully.

Validation used fake model providers and local Docker; live Daytona credentials were unset. Downloaded compatibility corpora were not present in this checkout.

[brynary]

@swerner RE: "a three-second fake response against a one-second budget" -- Does this mean a three second test? Sounds slow.