feat(agent): form pipeline + collection prompt workflow@20261021

- merge main: workflow/20261021 (collection prompt adds step-5 mandatory
  time readback-confirm before routing; non-motor/0003 branch rework;
  accident_cause unified schema), updated eval suites (photo_skip/
  button_guidance rewritten, no_false trimmed to asr, new
  non_motor_vehicle + time_confirmation with form/reply assertions),
  runner gains form/formMatches/replyMatches expectations
- prompts/collection.md regenerated from 20261021; extraction prompt
  archived (contentExtract description, 6-key whitelist)
- collect: extraction call per turn (extraction.md + current time/form
  injected), merge follows 更新表单 semantics (meaningful values only,
  number/string equality, diff -> turn_patch); accident_fields now uses
  external form keys
- backend: form_update flows to ChatResult and SSE formUpdate event
  (stream emits prefix, FormUpdate, body)
- fields: FORM_FIELDS whitelist with strict type validation
- verified: 155 unit tests; golden evals on real model - photo_skip 4/4,
  no_false 2/2, photo_button_guidance 12/12, non_motor_vehicle 14/14,
  collection_skip 10/12, time_confirmation 1/2; remaining 3 failures are
  model sequence/literal-template adherence, recorded in plan appendix

src/agent/collect.py

@@ -20,12 +20,11 @@ import json
 import re
 
 from .fields import (
-    FieldCandidate,
-    FieldStatus,
-    validate_candidate,
+    FORM_FIELDS,
+    validate_form_update,
 )
 from .llm import LLMService
-from .prompts import COLLECTION, render_prompt
+from .prompts import COLLECTION, EXTRACTION, render_prompt
 from .state import AgentState
 from .transitions import COLLECT_STAGE, is_legal_transition
 
@@ -131,37 +130,76 @@ def _build_user_message(state: AgentState) -> str:
 
 
 
-def _extract_candidates(state: AgentState, body: str) -> list[FieldCandidate]:
-    """从模型正文中提取字段候选（计划作业 4-4）。
+# 采集阶段表单状态使用工作流的外部表单键（has_injury 等，见 fields.FORM_FIELDS）；
+# 内部键（ywrysw 等）属于 get_info 分组，作业 6 再建立映射。
 
-    参考版返回空列表：字段提取依赖模型的结构化输出，留给真实
-    LLM 接入后实现（同一回复内 JSON 解析优先；确需独立提取调用时，
-    记录原因/调用次数/延迟）。validate_candidate 的字段级校验
-    已由单元测试覆盖，此处先接通数据流。
+# 提取调用的系统提示标记（不以 <state> 开头，不会被误当对话回复解析）
+_EXTRACTION_MARKER = "【字段提取】"
+
+
+async def _extract_turn_fields(state: AgentState, llm: LLMService) -> dict:
+    """独立提取调用（计划作业 4-4：调用计数可审计——每轮在对话调用之外多一次）。
+
+    提示词为工作流「文本内容提取事故信息」原文（含当前时间与当前表单
+    注入），模型只输出 JSON；校验见 fields.validate_form_update。
+    提取失败不阻塞对话（返回 {}）。
     """
-    return []
+    system = (
+        _EXTRACTION_MARKER
+        + "\n按下列系统提示词执行提取，只输出 JSON 对象，不输出解释。\n\n"
+        + render_prompt(EXTRACTION, state)
+    )
+    try:
+        raw = await llm.complete(
+            system,
+            state.get("input_text", ""),
+            history=conversation_history(state),
+        )
+        data = json.loads(raw.strip())
+    except Exception:
+        return {}
+    if not isinstance(data, dict):
+        return {}
+    return validate_form_update(data)
 
 
-def _apply_candidates(
-    state: AgentState, candidates: list[FieldCandidate]
-) -> tuple[dict, dict, dict]:
-    """把校验通过的候选写入正式字段。
+def _has_meaningful_value(value) -> bool:
+    """对应「更新表单」的 hasMeaningfulValue：None/空串无效（0 有效）。"""
+    return value is not None and value != ""
 
-    返回 (accident_fields, confirm_status, turn_patch)。
-    只有 CONFIRMED 的候选才写入正式字段并进入本轮差量；
-    CANDIDATE/UNKNOWN 只更新 confirm_status，供后续核实节点消费
-    （计划作业 4-5 的三态语义）。
+
+def _is_same_value(old, new) -> bool:
+    """对应「更新表单」的 isSameValue：数字与数字字符串视为相等。"""
+    if old == new:
+        return True
+    if old is None or new is None:
+        return False
+    if old == "" or new == "":
+        return False
+    try:
+        return float(old) == float(new)
+    except (TypeError, ValueError):
+        return str(old).strip() == str(new).strip()
+
+
+def _merge_form(state: AgentState, update: dict) -> tuple[dict, dict]:
+    """按「更新表单」语义合并：白名单内、有效、有变化的字段进入差量。
+
+    返回 (合并后的完整表单, 本轮差量 turn_patch)。
     """
-    accident_fields = dict(state.get("accident_fields") or {})
-    confirm_status = dict(state.get("confirm_status") or {})
+    form = dict(state.get("accident_fields") or {})
     patch: dict = {}
-    for cand in candidates:
-        validated = validate_candidate(cand)
-        confirm_status[validated.key] = validated.status
-        if validated.status is FieldStatus.CONFIRMED:
-            accident_fields[validated.key] = validated.value
-            patch[validated.key] = validated.value
-    return accident_fields, confirm_status, patch
+    for key in FORM_FIELDS:
+        if key not in update:
+            continue
+        value = update[key]
+        if not _has_meaningful_value(value):
+            continue
+        if _is_same_value(form.get(key), value):
+            continue
+        form[key] = value
+        patch[key] = value
+    return form, patch
 
 
 def _bookkeeping(state: AgentState) -> tuple[int, int]:
@@ -199,10 +237,9 @@ async def run_collect(state: AgentState, llm: LLMService) -> dict:
     )
 
 
-    # 4) 字段候选提取与三态落库（提取本身见 _extract_candidates）
-    accident_fields, confirm_status, patch = _apply_candidates(
-        state, _extract_candidates(state, raw)
-    )
+    # 4) 字段提取与表单差量（提取调用在对话调用之外，见 _extract_turn_fields）
+    extracted = await _extract_turn_fields(state, llm)
+    accident_fields, patch = _merge_form(state, extracted)
     no_reply_count, clarify_count = _bookkeeping(state)
 
     # 5) 状态增量：response_text 保留完整前缀文本，API 层负责解析；
@@ -218,7 +255,6 @@ async def run_collect(state: AgentState, llm: LLMService) -> dict:
         "stage_code": stage,
         "raw_status_code": stage,
         "accident_fields": accident_fields,
-        "confirm_status": confirm_status,
         "turn_patch": patch,
         "dialogue": dialogue,
     }