diff --git a/app/harness/validate.py b/app/harness/validate.py index 0e2adf7..4f35a43 100644 --- a/app/harness/validate.py +++ b/app/harness/validate.py @@ -809,7 +809,8 @@ class GateSpec: target_file: 解析后生效 skill 文件名(候选物化写此文件)。 candidate_content: 候选 skill 全文。 base_skill_content: 基线侧生效 skill 全文(skill_hash 作缓存键)。 - units: 阶梯序单元列表(已排除案例单元、截断 gate_n_max)。 + units: 阶梯序单元元组(已排除案例单元、截断 gate_n_max);元组,装配后 + 不可变,防 spec.units 与 run.slots 漂移。 gate_run_prefix: run_id 前缀,必须含 "_gate_"(防泄露过滤依赖)。 """ @@ -817,7 +818,7 @@ class GateSpec: target_file: str candidate_content: str base_skill_content: str - units: list[QuestionUnit] + units: tuple[QuestionUnit, ...] gate_run_prefix: str @@ -884,6 +885,10 @@ def _advance_prefix(run: _GateRun, params: GateParams) -> None: 瞬间返回、cand 臂必新鲜跑,两臂延迟不对称,若 cand 延迟与对错相关,早到翻转 对系统性偏向 W 型 → e-值虚高假接受。前缀消费把判定顺序钉回预声明阶梯序, anytime-valid 无条件成立;INFRA 单元视为"已解决(剔除)"不阻塞前缀。 + + 契约:全部单元被剔除时 verdict 保持 None、frozen 保持 False,由调度编排层 + (Task 3 的 validate_skills_concurrent)检测 verdict None 并 raise + RuntimeError;本函数不负责该终态。 """ while not run.frozen and run.prefix_ptr < len(run.slots): slot = run.slots[run.prefix_ptr] @@ -903,7 +908,9 @@ def _advance_prefix(run: _GateRun, params: GateParams) -> None: run.frozen = True continue uid = slot.unit.unit_id - assert slot.base is not None and slot.cand_per_q is not None + assert slot.base is not None and slot.cand_per_q is not None, ( + f"slot 未配齐即被消费: unit={slot.unit.unit_id}" + ) c_units = unit_correctness_view([slot.unit], slot.cand_per_q) pair_result = pair_block({uid: slot.base}, c_units, [uid]) run.candidate_per_q.update(slot.cand_per_q) diff --git a/tests/unit/test_gate_prefix.py b/tests/unit/test_gate_prefix.py index bf68ad6..8b6749b 100644 --- a/tests/unit/test_gate_prefix.py +++ b/tests/unit/test_gate_prefix.py @@ -33,7 +33,7 @@ def _mk_run(n_units: int) -> _GateRun: target_file="action-reasoning.md", candidate_content="cand", base_skill_content="base", - units=[_mk_unit(f"q{i}") for i in range(n_units)], + units=tuple(_mk_unit(f"q{i}") for i in range(n_units)), gate_run_prefix="r_e1_s0_gate_action-reasoning", ) return _GateRun.from_spec(spec) @@ -119,6 +119,48 @@ def test_infra_unit_skipped_not_counted() -> None: assert run.verdict is not None and run.verdict.decision == "accept_provisional" +def test_all_infra_leaves_verdict_none() -> None: + """全部单元被剔除:verdict 保持 None、frozen 保持 False——终态由调度编排层 + (Task 3 的 validate_skills_concurrent)检测 verdict None 并 raise,本函数不管。""" + run = _mk_run(3) + for i in range(3): + run.slots[i].base_infra = True + run.slots[i].cand_per_q = {f"q{i}": True} + _advance_prefix(run, _PARAMS) + assert run.verdict is None and not run.frozen and run.n_excluded == 3 and run.prefix_ptr == 3 + + +def test_repeated_calls_are_idempotent() -> None: + """部分前缀消费后重复调用不改变状态;补齐剩余单元后再调用正常推进。""" + run = _mk_run(4) + for i in (0, 1): + run.slots[i].base = False + run.slots[i].cand_per_q = {f"q{i}": True} + _advance_prefix(run, _PARAMS) + snapshot = (run.n_used, run.w, run.l, run.prefix_ptr, len(run.evidence_rows)) + _advance_prefix(run, _PARAMS) + _advance_prefix(run, _PARAMS) + assert (run.n_used, run.w, run.l, run.prefix_ptr, len(run.evidence_rows)) == snapshot + for i in (2, 3): + run.slots[i].base = False + run.slots[i].cand_per_q = {f"q{i}": True} + _advance_prefix(run, _PARAMS) + assert run.n_used > snapshot[0] and run.prefix_ptr > snapshot[3] + + +def test_frozen_run_call_is_noop() -> None: + """frozen 后再调用是 no-op:不再消费已配齐单元、证据不再追加。""" + run = _mk_run(12) + for i in range(12): + run.slots[i].base = False + run.slots[i].cand_per_q = {f"q{i}": True} + _advance_prefix(run, _PARAMS) + assert run.frozen and run.n_used == 7 + n_evidence = len(run.evidence_rows) + _advance_prefix(run, _PARAMS) + assert run.n_used == 7 and len(run.evidence_rows) == n_evidence + + def test_ties_hit_futility_early_and_freeze() -> None: """全打平(无翻转对)时 futility 出口尽早触发并冻结——早停语义(数值:W=L=0 时乐观 E = E(n_remaining, 0),n 小易