fix(scene): bind release gate to current readiness

2026-10-06 13:06:11 +08:00
parent a051131d3c
commit a601bc46a9
4 changed files with 112 additions and 1 deletions
@@ -681,6 +681,9 @@ class DraftingWorkspaceStore:
        summary.update({
            "source_applicability_verified": readiness.get("source_applicability_verified") is True,
            "published_prompt_alignment_verified": readiness.get("published_prompt_alignment_verified") is True,
+            # This is derived from the complete server-side readiness check.
+            # No edit, runtime, or API request field is accepted as a shortcut.
+            "release_readiness_prerequisites_satisfied": readiness.get("prerequisites_satisfied") is True,
            "release_readiness": readiness,
        })
        return summary
@@ -30,6 +30,8 @@ def scene_template_release_gate(
 ) -> dict:
    """Recompute release readiness from bound evidence, never cached approval flags."""
    evidence = model_evidence if isinstance(model_evidence, dict) else {}
+    readiness = evidence.get("release_readiness")
+    readiness = readiness if isinstance(readiness, dict) else {}
    blockers: list[dict[str, str]] = []

    def block(code: str, message: str) -> None:
@@ -89,6 +91,19 @@ def scene_template_release_gate(
        block("RELEASE_SOURCE_APPLICABILITY", "场景法源适用性尚未形成独立可核验结论。")
    if evidence.get("published_prompt_alignment_verified") is not True:
        block("RELEASE_RUNTIME_PROMPT_ALIGNMENT", "评测 Prompt 与拟发布运行时 Prompt 尚未完成哈希对齐。")
+    readiness_prerequisites_satisfied = (
+        evidence.get("release_readiness_prerequisites_satisfied") is True
+        and readiness.get("schema_version") == "scene-release-readiness.v1"
+        and readiness.get("prerequisites_satisfied") is True
+        and readiness.get("issues") == []
+        and readiness.get("source_applicability_verified") is True
+        and readiness.get("published_prompt_alignment_verified") is True
+        and readiness.get("missing_draft_variants") == []
+        and readiness.get("missing_review_variants") == []
+        and readiness.get("missing_holdout_draft_variants") == []
+    )
+    if not readiness_prerequisites_satisfied:
+        block("RELEASE_READINESS_PREREQUISITES", "服务端当前发布就绪条件尚未全部满足，历史 Prompt 覆盖不能替代当前运行证据。")

    bound_payload = {
        "policy_version": RELEASE_GATE_POLICY_VERSION,
@@ -103,6 +118,8 @@ def scene_template_release_gate(
            for group in complete_review],
        "source_applicability_verified": evidence.get("source_applicability_verified") is True,
        "published_prompt_alignment_verified": evidence.get("published_prompt_alignment_verified") is True,
+        "release_readiness_evidence_sha256": draft_digest(readiness) if readiness else None,
+        "release_readiness_prerequisites_satisfied": readiness_prerequisites_satisfied,
        "blocker_codes": [item["code"] for item in blockers],
    }
    return {
@@ -463,7 +463,9 @@ def test_scene_release_gate_recomputes_all_evidence_and_ignores_cached_approval_
    )
    assert passed["policy_version"] == RELEASE_GATE_POLICY_VERSION
    assert not passed["passed"] and not passed["authorizes_publication"]
-    assert {item["code"] for item in passed["blockers"]} == {"RELEASE_HOLDOUT_EVIDENCE_MISSING"}
+    assert {item["code"] for item in passed["blockers"]} == {
+        "RELEASE_HOLDOUT_EVIDENCE_MISSING", "RELEASE_READINESS_PREREQUISITES",
+    }
    assert len(passed["evidence_sha256"]) == 64

    stale = scene_template_release_gate(
@@ -477,6 +479,93 @@ def test_scene_release_gate_recomputes_all_evidence_and_ignores_cached_approval_
    })


+def test_final_publication_validation_rejects_complete_coverage_from_stale_prompts(monkeypatch):
+    from yuqei_contract_api.template_evaluation_evidence import (
+        POLICY_VERSION,
+        VARIANTS,
+        deterministic_scene_evidence,
+        scene_template_release_gate,
+    )
+
+    monkeypatch.setattr(
+        "yuqei_contract_api.holdout_scene_evidence_registry.verify_scene_holdout_evidence",
+        lambda evidence, *, current_template_content_sha256: True,
+    )
+    repository = InMemoryContractRepository(contracts=[])
+    template = repository.get_template(SCENE)
+    draft = repository.get_template_edit_draft(SCENE)
+    deterministic = deterministic_scene_evidence(template, draft.body_content)
+    stale_draft_prompt = {
+        "name": "contract_drafting_v1",
+        "version": "stale-version",
+        "prompt_sha256": "c" * 64,
+    }
+    stale_review_prompt = {
+        "name": "contract_review_transport",
+        "version": "stale-version",
+        "content_sha256": "d" * 64,
+    }
+    stale_readiness = {
+        "schema_version": "scene-release-readiness.v1",
+        "issues": [
+            {"code": "SCENE_DRAFT_COVERAGE_PENDING", "message": "旧 Prompt 起草覆盖不计入当前运行版本。"},
+            {"code": "SCENE_REVIEW_COVERAGE_PENDING", "message": "旧 Prompt 审查覆盖不计入当前运行版本。"},
+        ],
+        "source_applicability_verified": True,
+        "published_prompt_alignment_verified": True,
+        "missing_draft_variants": list(VARIANTS),
+        "missing_review_variants": list(VARIANTS),
+        "missing_holdout_draft_variants": [],
+        "prerequisites_satisfied": False,
+    }
+    evidence = {
+        "policy_version": POLICY_VERSION,
+        "template_content_sha256": deterministic["template_content_sha256"],
+        "current_records": len(VARIANTS),
+        "failed_or_unverified_records": 0,
+        "truncated": False,
+        "coverage_by_prompt": [{
+            "prompt": stale_draft_prompt,
+            "passed_variants": list(VARIANTS),
+            "missing_variants": [],
+        }],
+        "review_coverage_by_prompt": [{
+            "draft_prompt": stale_draft_prompt,
+            "prompt": stale_review_prompt,
+            "passed_variants": list(VARIANTS),
+            "missing_variants": [],
+        }],
+        "holdout_registry_evidence": {
+            "evidence_sha256": "e" * 64,
+            "current_template_match": True,
+            "missing_or_failed_variants": [],
+        },
+        "source_applicability_verified": True,
+        "published_prompt_alignment_verified": True,
+        # A cached or forged top-level claim cannot override server readiness.
+        "release_readiness_prerequisites_satisfied": True,
+        "release_readiness": stale_readiness,
+        "authorizes_publication": True,
+    }
+
+    gate = scene_template_release_gate(
+        template_content_sha256=deterministic["template_content_sha256"],
+        deterministic=deterministic,
+        model_evidence=evidence,
+    )
+    assert {item["code"] for item in gate["blockers"]} == {"RELEASE_READINESS_PREREQUISITES"}
+    assert gate["release_readiness_prerequisites_satisfied"] is False
+    assert not gate["passed"] and not gate["authorizes_publication"]
+
+    repository.get_template_evaluation_summary = lambda template_id: deepcopy(evidence)
+    validation = repository.validate_template_before_publish(SCENE)
+    assert not validation.valid
+    assert validation.evaluation_summary["release_gate"]["blocker_codes"] == [
+        "RELEASE_READINESS_PREREQUISITES"
+    ]
+    assert "SCENE_EVALUATION_PENDING" in {issue.code for issue in validation.issues}
+
+
 def test_evidence_recomputed_against_payload_not_cached_flags(stores):
    _, trial = stores
    workspace_id, _, _ = attach_stubbed_evaluation(trial)
@@ -106,6 +106,7 @@ def test_summary_top_level_release_flags_follow_only_verified_server_readiness(m
    summary = trial.template_evaluation_summary(SCENE, tenant="default")
    assert summary["source_applicability_verified"] is True
    assert summary["published_prompt_alignment_verified"] is True
+    assert summary["release_readiness_prerequisites_satisfied"] is summary["release_readiness"]["prerequisites_satisfied"]
    assert summary["source_applicability_verified"] is summary["release_readiness"]["source_applicability_verified"]
    assert summary["published_prompt_alignment_verified"] is summary["release_readiness"]["published_prompt_alignment_verified"]

@@ -122,6 +123,7 @@ def test_summary_top_level_release_flags_reject_cached_claims_and_bad_runtime_ha
    summary = trial.template_evaluation_summary(SCENE, tenant="default")
    assert summary["source_applicability_verified"] is False
    assert summary["published_prompt_alignment_verified"] is False
+    assert summary["release_readiness_prerequisites_satisfied"] is False
    assert "SCENE_RUNTIME_UNAVAILABLE" in codes(summary["release_readiness"])