Add minimal orchestrate mode with isolated scout/patch workspaces

[mmprotest]

Motivation

• Provide a small, deterministic outer loop to run multiple constrained Villani worker invocations in isolated worktrees to improve long-horizon task performance while avoiding a full multi-agent/chat redesign.
• Enforce that the orchestrator owns state, verification, scoring, and merge decisions and that workers perform only bounded, isolated repo work.

Description

• Add a new villani_code.orchestrate package implementing a simple round loop (DISCOVER → PLAN → EXECUTE → VERIFY → MERGE) with compact orchestration state saved to artifacts/orchestrate/<run_id>/state.json, and emit per-round artifacts plus final_report.json and summary.md (files: orchestrator.py, state.py, prompts.py, worker.py, worktree.py, verify.py, merge.py, __init__.py).
• Wire a new CLI command villani-code orchestrate "<task>" with the requested flags and passthrough for --provider/--model/--base-url/--api-key, and defaults: --workers=3, --scout-workers=3, --patch-workers=2, --rounds=3, --worker-timeout=300, --keep-worktrees and optional --verify-command. (villani_code/cli.py).
• Implement isolated workspaces via WorkspaceManager that prefers git worktree and falls back to copying a temp repo, enforces dirty-checks for scouts, and supports cleanup honoring --keep-worktrees (worktree.py).
• Implement a forgiving worker adapter and parser that runs the normal villani_code runner for workers, extracts a WORKER_REPORT_JSON block when present, and builds a resilient fallback report when parsing fails or a worker times out (worker.py, prompts.py).
• Add verification detection and execution helpers that reuse the existing benchmark verifier where possible, and a conservative candidate scoring and guarded-merge flow that only applies changes to the original repo after verification passes (verify.py, merge.py, orchestrator.py).

Testing

• Added unit and integration-like tests: tests/test_orchestrate_state.py, tests/test_orchestrate_worker.py, tests/test_orchestrate_merge.py, tests/test_orchestrate_worktree.py, and tests/test_orchestrate_integration.py to cover state serialization, report parsing, scoring, worktree cleanup, and artifact generation.
• Ran: pytest -q tests/test_orchestrate_state.py tests/test_orchestrate_worker.py tests/test_orchestrate_merge.py tests/test_orchestrate_worktree.py tests/test_orchestrate_integration.py tests/test_init_cli.py tests/test_cli_debug_flags.py, which passed (10 passed, 3 warnings).

---

Codex Task