fix(soak): 给 plan 与 execute 补上完成通路，它们原本必然跑满预算

打真实模型跑通一个完整任务时发现的：这两个阶段的工具集里没有带完成标记的工具，解释器
也从不产出最终回答，于是模型在第 4 步写完产出物之后还在继续读文档，直到撞上步数上限。
按原来的 50/50/16 算，二十个任务要两千三百多次调用，而且整批的停止原因会全是
step_budget，别的什么都压不出来。

解释器加一条最终回答支路（{"final_answer": "..."}），plan 与 execute 的提示词写死收尾
动作；summarize 不变，仍然只能靠 submit_finding 结束。这样三条完成路径同时在场：模型
自报最终回答、agent 调用带完成标记的工具、环境报告完成（AppWorld 那路），它们的可信度
各不相同，压测正需要这个对照。

预算随之下调到 20/25/16——实测每阶段真实用 5 到 7 步，留了一倍余量。原来那个 50 的来历
是 gov-auditor.yaml 的 turns 上限，但那边靠编排校验产物落盘来收阶段，跑满 turns 无所谓；
这里靠模型自己收尾，上限定高只会让它在产出物写完之后接着白烧。

实测：plan 7 步 agent_finished、execute 6 步 agent_finished、summarize 5 步
task_completed，一个任务 18 次调用。

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

tools/soak/tests/test_govdoc_scenario.py

@@ -21,7 +21,7 @@ from typing import TYPE_CHECKING
 
 import pytest
 
-from polyloop.ports import Action, InvalidDecision
+from polyloop.ports import Action, FinalAnswer, InvalidDecision
 from polyloop.tools import ToolRegistry
 from polyloop.types import ModelReply, ReplayPolicy
 from tools.soak.scenarios.govdoc import (
@@ -217,7 +217,7 @@ def test_contract_4_action_carries_text():
 )
 def test_contract_5_parse_never_raises(content: str):
     parsed = GovDocParser().parse(_reply(content))
-    assert isinstance(parsed.decision, Action | InvalidDecision)
+    assert isinstance(parsed.decision, Action | FinalAnswer | InvalidDecision)
 
 
 def test_tolerates_code_fences():
@@ -264,6 +264,74 @@ def test_five_parse_failures_get_five_explanations():
     assert "arguments" in explanations["arguments 不是对象"]
 
 
+def test_final_answer_branch():
+    parsed = GovDocParser().parse(_reply('{"final_answer": "已写出 plan.md，列了 3 处候选证据。"}'))
+    assert isinstance(parsed.decision, FinalAnswer)
+    assert parsed.decision.text == "已写出 plan.md，列了 3 处候选证据。"
+    assert len(parsed.history_text) <= len(
+        '{"final_answer": "已写出 plan.md，列了 3 处候选证据。"}'
+    )
+
+
+def test_final_answer_tolerates_code_fences():
+    inner = '{"final_answer": "已写出 evidence.md，3 条证据。"}'
+    for fenced in (
+        f"```json\n{inner}\n```",
+        f"```\n{inner}\n```",
+        f"这一阶段做完了：\n```json\n{inner}\n```\n",
+    ):
+        parsed = GovDocParser().parse(_reply(fenced))
+        assert isinstance(parsed.decision, FinalAnswer), fenced
+        assert parsed.decision.text == "已写出 evidence.md，3 条证据。"
+
+
+def test_tool_wins_when_both_keys_are_present():
+    """一边调工具一边宣布做完时以工具为准：按 final_answer 收尾会把那次调用整个丢掉。"""
+    parsed = GovDocParser().parse(
+        _reply(
+            '{"tool": "write_note", "arguments": {"filename": "plan.md", "content": "x"},'
+            ' "final_answer": "我写完了"}'
+        )
+    )
+    assert isinstance(parsed.decision, Action)
+    assert parsed.decision.tool_call is not None
+    assert parsed.decision.tool_call.name == "write_note"
+
+
+def test_flattened_arguments_do_not_swallow_final_answer():
+    parsed = GovDocParser().parse(
+        _reply(
+            '{"tool": "write_note", "filename": "plan.md", "content": "x", "final_answer": "完"}'
+        )
+    )
+    assert isinstance(parsed.decision, Action)
+    assert parsed.decision.tool_call is not None
+    assert parsed.decision.tool_call.arguments == {"filename": "plan.md", "content": "x"}
+
+
+@pytest.mark.parametrize(
+    "content",
+    [
+        '{"answer": "我做完了"}',
+        '{"arguments": {"path": "tender.md"}}',
+        '{"final_answer": ""}',
+        '{"final_answer": "   "}',
+        "{}",
+    ],
+)
+def test_neither_key_is_still_invalid(content: str):
+    decision = GovDocParser().parse(_reply(content)).decision
+    assert isinstance(decision, InvalidDecision)
+    assert decision.explanation.strip()
+
+
+def test_missing_key_explanation_names_both_shapes():
+    decision = GovDocParser().parse(_reply('{"note": "x"}')).decision
+    assert isinstance(decision, InvalidDecision)
+    assert "tool" in decision.explanation
+    assert "final_answer" in decision.explanation
+
+
 def test_parser_reports_its_parameters():
     assert GovDocParser().parameters() == {"kind": "govdoc_json_tool_call"}
 
@@ -478,7 +546,11 @@ def test_all_three_phases_assemble(tmp_path: Path):
         assert request.injections == {}
 
 
-def test_budgets_follow_gov_auditor_turn_limits(tmp_path: Path):
+def test_budgets_are_the_measured_step_counts(tmp_path: Path):
+    """20 / 25 / 16 是实测出来的，不是照抄 gov-auditor.yaml 的 50 / 50 / 16。
+
+    断言具体数字是因为这三个数直接决定一次全量压测的调用量，改动必须是有意的。
+    """
     task = _sample_task()
     budgets = {
         phase: build_run_request(
@@ -490,12 +562,35 @@ def test_budgets_follow_gov_auditor_turn_limits(tmp_path: Path):
         ).budget
         for phase in ("plan", "execute", "summarize")
     }
-    assert (budgets["plan"].max_steps, budgets["plan"].max_actions) == (50, 50)
-    assert (budgets["execute"].max_steps, budgets["execute"].max_actions) == (50, 50)
+    assert (budgets["plan"].max_steps, budgets["plan"].max_actions) == (20, 20)
+    assert (budgets["execute"].max_steps, budgets["execute"].max_actions) == (25, 25)
     assert (budgets["summarize"].max_steps, budgets["summarize"].max_actions) == (16, 16)
     for budget in budgets.values():
         assert budget.max_consecutive_parse_failures == 3
         assert budget.max_prompt_chars == 400_000
+    # 一次全量（20 个任务 × 3 个阶段）的步数上限。这一条守的是额度，不是行为。
+    assert sum(budget.max_steps for budget in budgets.values()) * 20 == 1220
+
+
+def test_plan_and_execute_prompts_spell_out_the_final_answer_closing(tmp_path: Path):
+    """plan 与 execute 的工具集里没有带 completes_run 的工具，收尾只能靠最终回答。
+
+    提示词不写清楚这一步，这两个阶段就必然跑满预算——实测过，停止原因全是 step_budget。
+    """
+    task = _sample_task()
+    registry = _tools(tmp_path).registry()
+    for phase in ("plan", "execute"):
+        narrowed = registry.restrict_to(PHASE_TOOLS[phase])
+        system = build_context(task=task, phase=phase, tools=narrowed).run_level[0].content[0].text
+        assert '{"final_answer":' in system
+        assert not any(narrowed.spec_for(name).completes_run for name in narrowed.names())
+    # summarize 不变：它靠 submit_finding 这条提交型完成通路结束，不给最终回答这条路。
+    summarize = registry.restrict_to(PHASE_TOOLS["summarize"])
+    system = (
+        build_context(task=task, phase="summarize", tools=summarize).run_level[0].content[0].text
+    )
+    assert "final_answer" not in system
+    assert any(summarize.spec_for(name).completes_run for name in summarize.names())
 
 
 def test_context_carries_tools_but_not_the_document_body(tmp_path: Path):