Add an explicit termination reason to the finish tool

[mdev34-lab]

Problem

The finish tool currently communicates that an agent has stopped, but the runtime cannot reliably distinguish why it stopped. This becomes increasingly important if Review/subagent execution is moved onto the background path.

Proposal

Add a required reason parameter to finish, alongside the existing description:

finish({
  reason: "success" | "subagent_wait" | "failure",
  description: string,
})

Semantics

• success: the agent completed its assigned work and has a usable result.
• subagent_wait: the agent is done with its current turn because progress depends on another subagent; the scheduler/orchestrator should handle the dependency.
• failure: the agent could not complete its assigned work successfully.

Requirements

• Make reason required rather than optional.
• Preserve the existing description/result payload.
• Propagate the reason through the existing finish/result handling instead of inferring it from prose.
• Update the finish tool schema, runtime types, dispatch/termination handling, and relevant tests.
• Ensure background and foreground subagents expose the same termination contract.
• Do not make subagent_wait equivalent to fire-and-forget; the existing orchestration/scheduler semantics should remain explicit.
• Audit existing callers/tests for the new required field and update them.
• Keep the change focused; do not redesign the entire subagent scheduler as part of this issue.

Motivation

An explicit termination reason gives the runtime a typed state transition instead of forcing it to infer termination semantics from the model's text. This is particularly useful for background Review/subagent execution, recovery from non-terminal model behavior, orchestration between subagents, and future observability.

This issue is related to the Review stagnation recovery work in #171 and the broader finish/liveness work in #140, but should remain independently implementable.