fix(rectification): surface dropped probes and filter tools by the decision

Silent unrenderable discriminators, a missing question-contract golden, and a always-on tool table were hiding fail-closed drops behind the prompt wall.

Co-authored-by: Cursor <cursoragent@cursor.com>

scripts/rectification/probe_question_contract.py

@@ -1,12 +1,15 @@
 """Shared probe → choice-card contract for Python event probes.
 
 Must stay aligned with frontend/src/lib/rectification-agentic/v9/probe-question-contract.ts.
+Canonical bytes live in contracts/probe-question-v1.json.
 """
 
 from __future__ import annotations
 
+import json
 import re
-from typing import Any, Sequence
+from pathlib import Path
+from typing import Any, Sequence, TypedDict
 
 QUESTION_CONTRACT_VERSION = "probe-question-v1"
 ANSWER_CLASSES = ("yes", "weak_yes", "no", "unsure")
@@ -24,8 +27,32 @@ QUALITY_STYLE_OPTIONS: tuple[dict[str, str], ...] = (
 )
 VARGA_NONE_STYLE_OPTION = {"label": "都不是这些特质", "answer_class": "no"}
 UNSURE_STYLE_OPTION = {"label": "这段记不清楚", "answer_class": "unsure"}
-_FORBIDDEN = ("外貌", "体质", "胎记", "疤痕", "伤疤", "身高", "体型")
+FORBIDDEN_COPY_TOKENS = ("外貌", "体质", "胎记", "疤痕", "伤疤", "身高", "体型")
+LABEL_MIN = 4
+LABEL_MAX = 80
+_FORBIDDEN = FORBIDDEN_COPY_TOKENS
 _CLOCK = re.compile(r"(?:[01]?\d|2[0-3]):[0-5]\d")
+_ANSWER_CLASS_SUFFIX = re.compile(r"·(?:yes|weak_yes|no|unsure)$")
+_CONTRACT_PATH = Path(__file__).resolve().parents[2] / "contracts" / "probe-question-v1.json"
+
+
+class StyleOptionsOk(TypedDict):
+    ok: bool
+    options: list[dict[str, str]]
+
+
+class StyleOptionsErr(TypedDict):
+    ok: bool
+    reason: str
+
+
+class ProbeRenderOk(TypedDict):
+    ok: bool
+
+
+class ProbeRenderErr(TypedDict):
+    ok: bool
+    reason: str
 
 
 def probe_question_kind(value: Any) -> str:
@@ -34,7 +61,40 @@ def probe_question_kind(value: Any) -> str:
     return "existence"
 
 
-def clipped_probe_label(value: Any, minimum: int = 4, maximum: int = 80) -> str | None:
+def question_contract_version_is_compatible(value: Any) -> bool:
+    if value is None:
+        return True
+    if isinstance(value, str):
+        return value == QUESTION_CONTRACT_VERSION
+    if not isinstance(value, dict):
+        return False
+    version = value.get("version") or value.get("question_contract_version")
+    return version is None or version == QUESTION_CONTRACT_VERSION
+
+
+def probe_question_contract_payload() -> dict[str, Any]:
+    return {
+        "version": QUESTION_CONTRACT_VERSION,
+        "answer_classes": list(ANSWER_CLASSES),
+        "label_min": LABEL_MIN,
+        "label_max": LABEL_MAX,
+        "forbidden_copy_tokens": list(FORBIDDEN_COPY_TOKENS),
+        "existence_style_options": [dict(item) for item in EXISTENCE_STYLE_OPTIONS],
+        "quality_style_options": [dict(item) for item in QUALITY_STYLE_OPTIONS],
+        "varga_none_style_option": dict(VARGA_NONE_STYLE_OPTION),
+        "unsure_style_option": dict(UNSURE_STYLE_OPTION),
+    }
+
+
+def canonical_probe_question_contract_json() -> str:
+    return json.dumps(probe_question_contract_payload(), ensure_ascii=False, indent=2) + "\n"
+
+
+def load_probe_question_contract_golden() -> str:
+    return _CONTRACT_PATH.read_text(encoding="utf-8")
+
+
+def clipped_probe_label(value: Any, minimum: int = LABEL_MIN, maximum: int = LABEL_MAX) -> str | None:
     if not isinstance(value, str):
         return None
     text = " ".join(value.split())
@@ -45,13 +105,29 @@ def clipped_probe_label(value: Any, minimum: int = 4, maximum: int = 80) -> str
     return text
 
 
-def _incoming_option(row: Any) -> dict[str, str] | None:
+def _label_reject_reason(value: Any) -> str | None:
+    if not isinstance(value, str):
+        return "empty"
+    text = " ".join(value.split())
+    if len(text) < LABEL_MIN or len(text) > LABEL_MAX:
+        return "label_length"
+    if any(token in text for token in _FORBIDDEN) or _CLOCK.search(text):
+        return "forbidden_copy"
+    return None
+
+
+def _incoming_option(row: Any) -> dict[str, str] | dict[str, str] | None:
     if not isinstance(row, dict):
         return None
     answer = row.get("answer_class") or row.get("answerClass")
-    label = clipped_probe_label(row.get("label"))
-    if not label or answer not in ANSWER_CLASSES:
+    reject = _label_reject_reason(row.get("label"))
+    if reject == "empty" or answer not in ANSWER_CLASSES:
         return None
+    if reject:
+        return {"reason": reject}
+    label = clipped_probe_label(row.get("label"))
+    if not label:
+        return {"reason": "not_renderable"}
     payload = {"label": label, "answer_class": str(answer)}
     sign = row.get("sign")
     if isinstance(sign, str) and sign.strip():
@@ -62,9 +138,18 @@ def _incoming_option(row: Any) -> dict[str, str] | None:
 def complete_style_options(
     choice_kind: Any,
     style_options: Sequence[Any] | None = None,
-) -> list[dict[str, str]] | None:
+) -> StyleOptionsOk | StyleOptionsErr:
     kind = probe_question_kind(choice_kind)
-    incoming = [item for item in (_incoming_option(row) for row in (style_options or [])) if item]
+    incoming_reason: str | None = None
+    incoming: list[dict[str, str]] = []
+    for row in style_options or []:
+        parsed = _incoming_option(row)
+        if not parsed:
+            continue
+        if "reason" in parsed and "label" not in parsed:
+            incoming_reason = incoming_reason or str(parsed["reason"])
+            continue
+        incoming.append(parsed)
     by_class: dict[str, dict[str, str]] = {}
     if kind == "varga_style":
         for option in incoming:
@@ -72,10 +157,10 @@ def complete_style_options(
         by_class.setdefault("unsure", dict(UNSURE_STYLE_OPTION))
         scoring = [item for item in ANSWER_CLASSES if item != "unsure" and item in by_class]
         if len(scoring) < 2:
-            return None
+            return {"ok": False, "reason": incoming_reason or "varga_insufficient_scoring"}
         by_class.setdefault("no", dict(VARGA_NONE_STYLE_OPTION))
         if "yes" not in by_class or "weak_yes" not in by_class:
-            return None
+            return {"ok": False, "reason": "varga_missing_weak_yes"}
     else:
         catalog = QUALITY_STYLE_OPTIONS if kind == "event_quality" else EXISTENCE_STYLE_OPTIONS
         for option in catalog:
@@ -84,32 +169,50 @@ def complete_style_options(
             by_class[option["answer_class"]] = option
     ordered: list[dict[str, str]] = []
     seen: set[str] = set()
-    for answer_class in ANSWER_CLASSES:
+    for index, answer_class in enumerate(ANSWER_CLASSES):
         option = by_class.get(answer_class)
         if not option:
-            return None
+            return {"ok": False, "reason": "not_renderable"}
         label = option["label"]
         if label in seen and option.get("sign"):
             label = f"{label}（{option['sign']}）"
         if label in seen:
-            label = f"{label}·{answer_class}"
+            label = f"{option['label']}·{index + 1}"
         seen.add(label)
         ordered.append({**option, "label": label})
     labels = {item["label"] for item in ordered}
     classes = {item["answer_class"] for item in ordered}
-    if len(ordered) != 4 or labels != {item["label"] for item in ordered} or classes != set(ANSWER_CLASSES):
-        return None
-    if len(labels) != 4:
-        return None
-    return ordered
+    if (
+        len(ordered) != 4
+        or len(labels) != 4
+        or classes != set(ANSWER_CLASSES)
+        or any(_ANSWER_CLASS_SUFFIX.search(item["label"]) for item in ordered)
+    ):
+        return {"ok": False, "reason": "not_renderable"}
+    return {"ok": True, "options": ordered}
 
 
-def is_renderable_probe(probe: dict[str, Any]) -> bool:
+def completed_style_options(
+    choice_kind: Any,
+    style_options: Sequence[Any] | None = None,
+) -> list[dict[str, str]] | None:
+    result = complete_style_options(choice_kind, style_options)
+    if result.get("ok"):
+        return result.get("options")  # type: ignore[return-value]
+    return None
+
+
+def is_renderable_probe(probe: dict[str, Any]) -> ProbeRenderOk | ProbeRenderErr:
     gain = probe.get("information_gain")
     if not isinstance(gain, (int, float)) or gain <= 0:
-        return False
+        return {"ok": False, "reason": "zero_gain"}
     candidate_ids = probe.get("candidate_ids") or []
     outcomes = probe.get("expected_outcomes") or []
-    if len(candidate_ids) < 2 or len(outcomes) < 2:
-        return False
-    return complete_style_options(probe.get("choice_kind"), probe.get("style_options")) is not None
+    if len(candidate_ids) < 2:
+        return {"ok": False, "reason": "insufficient_candidates"}
+    if len(outcomes) < 2:
+        return {"ok": False, "reason": "insufficient_outcomes"}
+    completed = complete_style_options(probe.get("choice_kind"), probe.get("style_options"))
+    if not completed.get("ok"):
+        return {"ok": False, "reason": str(completed.get("reason") or "not_renderable")}
+    return {"ok": True}