fix(agent): reserve final step for response

[mikemikimike]

Summary

When an agent used every configured step for non-terminal tools, the loop emitted a synthetic max-steps reply without giving the model a chance to summarize completed work. This change reserves the final budgeted inference for reply or finish.

Changes

• Limit the final step's prompt and native tool payload to the terminal tools.
• Reject non-terminal calls during final-step parsing before any tool execution.
• Preserve the existing stalled/max-steps result if finalization cannot produce a terminal response.
• Add an AgentLoop regression test covering a tool step followed by a final reply.

Validation

• npm test -- --run src/agent/agent-loop.test.ts src/agent/step-executor.test.ts (64 passed)
• npm run lint (passed)
• git diff --check (passed)

The full test run reached 6,290 passing tests; 17 unrelated tests failed due to existing Windows/environment issues, including POSIX-only executable/path assumptions, temporary-directory teardown locks, a missing eval fixture, and platform-sensitive timing tests.