feat: 来源归因规则引擎（渠道/AI/搜索词/诚实状态/历史重算）

- constants 规则库 v1：AI/引擎/社交域名表、搜索词参数白名单、付费点击参数、
  应用内 UA 标记、CHANNELS 枚举；page_view 白名单开 nav_type/prerendering
- services/attribution.classify 纯函数：付费>AI>引擎>社交>referral 分层，
  UTM 证据叠加（combined/conflict/label_only），应用内 UA 叠加/兜底，
  引擎 TLD 国家 variant，搜索词白名单提取（Google 恒空不虚构）
- SourceClassification 派生表（迁移 0004）：first_touch/session_entry 双 scope，
  rule_version 可追溯
- ingest 增量 best-effort：首访客/新会话时分类落库（savepoint+静默，
  失败不影响采集）；validation 对 page_view referrer 完整 URL 参数清洗+截500
- recompute_attribution 命令：全量/按站幂等重算
- API：访客列表归因五字段+channel 筛选、详情 attribution 块（含最近5会话入口）、
  dashboard channels 聚合（sources 原口径保留）；TS VisitorListParams 加 channel
- 测试 97 → 124（引擎20 + 集成/接口7）

Co-Authored-By: Claude Code <noreply@anthropic.com>

admin/src/api/tracker/visitors.ts

@@ -132,6 +132,9 @@ export interface VisitorListParams {
   show_bot?: '1'
   /** visitor_id / last_ip 前缀模糊匹配 */
   search?: string
+  /** 渠道筛选（CHANNELS 枚举：natural_search/paid_search/ai_referral/social/in_app/referral/direct/unknown，
+   *  按访客 first_touch 分类过滤；无分类行的历史访客不属于任何渠道桶） */
+  channel?: string
   /** YYYY-MM-DD（非法值后端宽松忽略） */
   date_from?: string
   /** YYYY-MM-DD */

server/tracker/api/dashboard.py

@@ -4,7 +4,7 @@ from django.utils import timezone
 from rest_framework.views import APIView
 from rest_framework.response import Response
 from tracker.api.envelope import EnvelopeMixin, ok
-from tracker.models import DailyStat, Event, Site, VisitSession, Visitor
+from tracker.models import DailyStat, Event, Site, SourceClassification, VisitSession, Visitor
 from tracker.services.stats import daily_counts
 
 METRICS = ("pv", "uv", "sessions", "new_visitors")
@@ -31,6 +31,7 @@ class OverviewView(EnvelopeMixin, APIView):
         return ok({"today": counts,
                    "trend": self._trend(site, today, counts),
                    "sources": self._today_sources(site, today),
+                   "channels": self._today_channels(site, today),
                    "top_pages": self._today_pages(site, today)})
 
     @staticmethod
@@ -75,6 +76,22 @@ class OverviewView(EnvelopeMixin, APIView):
                          .annotate(n=Count("id"))
                          .order_by("-n", "referrer_domain")[:8]]   # 平局时次序跨请求稳定
 
+    def _today_channels(self, site, today):
+        """当天会话按来源渠道聚合（session_entry 分类；sources 的升级版，原始口径保留）。
+        从分类表正向聚合 + Count(distinct session)：无唯一约束下分类表偶发同键重复行
+        也不会把同一会话计两次。无分类行的会话（未重算的历史数据）不计入。"""
+        start, end = _day_bounds(today)
+        q = SourceClassification.objects.filter(
+            scope="session_entry",
+            session__visitor__is_bot=False,
+            session__started_at__gte=start, session__started_at__lt=end)
+        if site is not None:
+            q = q.filter(site=site)
+        return [{"channel": r["channel"] or "unknown", "sessions": r["n"]}
+                for r in q.values("channel")
+                         .annotate(n=Count("session", distinct=True))
+                         .order_by("-n", "channel")[:8]]   # 平局时次序跨请求稳定
+
     def _today_pages(self, site, today):
         start, end = _day_bounds(today)
         q = Event.objects.filter(is_bot=False, event_name="page_view",

server/tracker/api/visitors.py

@@ -13,7 +13,7 @@ from rest_framework.response import Response
 from rest_framework.views import APIView
 
 from tracker.api.envelope import EnvelopeMixin, ok
-from tracker.models import Event, RawBatch, Site, VisitSession, Visitor
+from tracker.models import Event, RawBatch, Site, SourceClassification, VisitSession, Visitor
 
 PAGE_SIZE = 20
 RAW_PAGE_SIZE = 10   # 原始批次 body 较大，页小一些
@@ -21,6 +21,7 @@ TIMELINE_DEFAULT_DAYS = 30
 TIMELINE_MAX_DAYS = 90
 TIMELINE_MAX_EVENTS = 1000
 ONLINE_WINDOW = timedelta(minutes=5)   # 与 dashboard active_now 同口径
+ATTR_SESSION_ENTRIES = 5               # 详情页归因块带出的最近会话入口分类数
 
 
 def _source(referrer_domain, utm):
@@ -34,8 +35,9 @@ def _iso(dt):
     return timezone.localtime(dt).isoformat() if dt else None
 
 
-def _row(visitor, last_session, online_cutoff):
+def _row(visitor, last_session, online_cutoff, ft=None):
-    """列表/详情共用行字段。last_source 依赖最近会话（由调用方批量取好传入，防 N+1）。"""
+    """列表/详情共用行字段。last_source 依赖最近会话（由调用方批量取好传入，防 N+1）。
+    channel 等归因五字段来自该访客 first_touch 分类（无分类行 → 空串，前端回落 first_source 展示）。"""
     return {
         "visitor_id": visitor.visitor_id,
         # 站点归属：visitor.site 需随行 select_related("site")，否则逐行多一条查询
@@ -51,10 +53,28 @@ def _row(visitor, last_session, online_cutoff):
         "first_source": _source(visitor.first_referrer_domain, visitor.first_utm),
         "last_source": _source(last_session.referrer_domain, last_session.utm)
                        if last_session else "直接访问",   # 无会话访客（仅采集异常场景）
+        "channel": ft.channel if ft else "",
+        "platform": ft.platform if ft else "",
+        "variant": ft.variant if ft else "",
+        "search_term": ft.search_term if ft else "",
+        "resolution": ft.resolution if ft else "",
         "online": visitor.last_seen_at > online_cutoff,   # Python 侧比对（last_seen_at 已随行取出）
     }
 
 
+def _classification(sc, with_session=False):
+    """SourceClassification → API dict（详情归因块用，含证据可信度与规则版本）。"""
+    if not sc:
+        return None
+    d = {"channel": sc.channel, "platform": sc.platform, "variant": sc.variant,
+         "search_term": sc.search_term, "evidence_kind": sc.evidence_kind,
+         "resolution": sc.resolution, "detail": sc.detail or {},
+         "rule_version": sc.rule_version, "computed_at": _iso(sc.computed_at)}
+    if with_session:
+        d["session_id"] = sc.session.session_id if sc.session_id else None
+    return d
+
+
 def _session_summary(s):
     # duration_ms 仅内部字段不展示——前端时长一律用 stay_ms（有效停留口径）
     return {"session_id": s.session_id,
@@ -96,6 +116,11 @@ def _page_visitors(qs, request):
     if p.get("search"):
         q = p["search"]
         qs = qs.filter(Q(visitor_id__istartswith=q) | Q(last_ip__istartswith=q))
+    if p.get("channel"):
+        # 渠道筛选（值=CHANNELS 枚举）：按 first_touch 分类内连接过滤。
+        # distinct 防御分类表脏重复行（无唯一约束，并发极端下可能同键两行）撑大 count。
+        qs = qs.filter(source_classifications__scope="first_touch",
+                       source_classifications__channel=p["channel"]).distinct()
     date_from = _parse_date(p.get("date_from"))
     date_to = _parse_date(p.get("date_to"))
     if date_from:
@@ -114,8 +139,12 @@ def _page_visitors(qs, request):
     visitors = list(page_qs[(page - 1) * PAGE_SIZE: page * PAGE_SIZE])
     latest = {s.pk: s for s in VisitSession.objects.filter(
         pk__in=[v.latest_session_pk for v in visitors if v.latest_session_pk])}
+    # 归因字段批量取：整页一次 IN 查询（同 latest 的做法），不做 select_related N+1
+    fts = {sc.visitor_id: sc for sc in SourceClassification.objects.filter(
+        visitor__in=[v.pk for v in visitors], scope="first_touch")}
     cutoff = timezone.now() - ONLINE_WINDOW
-    rows = [_row(v, latest.get(v.latest_session_pk), cutoff) for v in visitors]
+    rows = [_row(v, latest.get(v.latest_session_pk), cutoff, fts.get(v.pk))
+            for v in visitors]
     return rows, total, page
 
 
@@ -190,11 +219,20 @@ class VisitorDetailView(EnvelopeMixin, APIView):
         # 全量会话摘要（visitor,-started_at 索引）；M1 访客会话量有限不分页
         sessions = VisitSession.objects.filter(visitor=visitor).order_by("-started_at", "-pk")
         cutoff = timezone.now() - ONLINE_WINDOW
-        return ok({**_row(visitor, last_session, cutoff),
+        ft = (SourceClassification.objects.filter(visitor=visitor, scope="first_touch")
+              .first())
+        entries = (SourceClassification.objects
+                   .filter(visitor=visitor, scope="session_entry")
+                   .select_related("session")            # session_id 随行带出，免逐行查询
+                   .order_by("-session__started_at", "-pk")[:ATTR_SESSION_ENTRIES])
+        return ok({**_row(visitor, last_session, cutoff, ft),
                    "screen": visitor.screen, "timezone_name": visitor.timezone_name,
                    "last_ip": visitor.last_ip, "first_utm": visitor.first_utm or {},
                    "first_landing_path": visitor.first_landing_path,
                    "first_referrer_domain": visitor.first_referrer_domain,
+                   "attribution": {"first_touch": _classification(ft),
+                                   "session_entries": [_classification(e, with_session=True)
+                                                       for e in entries]},
                    "sessions": [_session_summary(s) for s in sessions]})
 
 

server/tracker/constants.py

@@ -11,9 +11,41 @@ URL_PARAM_DROP_PATTERNS = ("email", "phone", "mobile", "token", "key", "password
 SESSION_IDLE_MS = 30 * 60 * 1000
 BOT_UA_KEYWORDS = ("bot", "spider", "crawl", "slurp", "bingpreview", "headless", "lighthouse")
 EVENT_PROP_WHITELIST = {          # spec §9.1：未知字段拒绝。免费文本/任意属性不得入库
-    "page_view": {"referrer", "referrer_domain", "utm"},
+    "page_view": {"referrer", "referrer_domain", "utm", "nav_type", "prerendering"},
     "page_stay": {"active_ms", "path"},
     "sdk_loaded": {"sdk_version"},
 }
 MAX_PROP_DEPTH = 2
 MAX_PROP_JSON_BYTES = 4096
+
+# ---------------------------------------------------------------------------
+# 来源归因规则库 v1（tracker/services/attribution.py 消费）。改规则必须 bump 版本号
+# 并跑 recompute_attribution 重算历史——已落库分类带 rule_version 可追溯。
+# ---------------------------------------------------------------------------
+ATTRIBUTION_RULE_VERSION = "v1"
+AI_REFERRAL_DOMAINS = {  # AI 平台 → 显示名
+    "chatgpt.com": "ChatGPT", "perplexity.ai": "Perplexity",
+    "copilot.microsoft.com": "Copilot", "gemini.google.com": "Gemini",
+    "deepseek.com": "DeepSeek", "claude.ai": "Claude", "poe.com": "Poe", "you.com": "You",
+}
+SEARCH_ENGINE_DOMAINS = {  # 引擎后缀/域名 → 显示名（"google."/"yandex." 前缀匹配任意国家站 TLD）
+    "google.": "Google", "bing.com": "Bing", "baidu.com": "百度", "yandex.": "Yandex",
+    "duckduckgo.com": "DuckDuckGo", "yahoo.com": "Yahoo", "ecosia.org": "Ecosia",
+    "search.naver.com": "Naver", "search.brave.com": "Brave",
+}
+SEARCH_TERM_PARAMS = {  # 引擎 → referrer URL 里的搜索词参数名（白名单提取）
+    "Bing": ["q"], "百度": ["wd", "word"], "DuckDuckGo": ["q"], "Yandex": ["text"],
+    "Ecosia": ["q"], "Yahoo": ["p"], "Brave": ["q"], "Naver": ["query"],
+}  # 注：Google 不提供，绝不虚构
+PAID_CLICK_PARAMS = {"gclid": "Google Ads", "fbclid": "Meta Ads",
+                     "msclkid": "Bing Ads", "yclid": "Yandex Ads"}
+SOCIAL_DOMAINS = {
+    "t.co": "Twitter/X", "twitter.com": "Twitter/X", "x.com": "Twitter/X",
+    "facebook.com": "Facebook", "linkedin.com": "LinkedIn", "instagram.com": "Instagram",
+    "vk.com": "VK", "reddit.com": "Reddit", "pinterest.com": "Pinterest",
+    "telegram.me": "Telegram", "t.me": "Telegram",
+}
+INAPP_UA_MARKERS = {"MicroMessenger": "微信", "Line/": "Line", "FBAN": "Facebook",
+                    "FBAV": "Facebook", "Instagram": "Instagram", "WhatsApp": "WhatsApp"}
+CHANNELS = ("natural_search", "paid_search", "ai_referral", "social",
+            "in_app", "referral", "direct", "unknown")

server/tracker/management/commands/recompute_attribution.py

@@ -0,0 +1,88 @@
+"""历史来源分类全量重算（幂等可重复）。
+
+规则引擎升级（constants.ATTRIBUTION_RULE_VERSION bump）后执行：
+    python manage.py recompute_attribution            # 全部站点
+    python manage.py recompute_attribution --site site_xxx
+
+数据源约束：重算只能用"落库时保住的信号"——
+  - 访客首触：最早 page_view 的 properties.referrer/utm（无则档案字段兜底，
+    referrer_domain 合成 URL 后搜索词已丢失，诚实降级）+ first_landing_path；UA 未存档 → 空。
+  - 会话入口：referrer_domain 合成 URL + utm + landing_path。
+逐条 upsert（first_touch 键 site+visitor+scope；session_entry 键 site+session+scope），
+跑两遍计数不变。
+"""
+from django.core.management.base import BaseCommand
+
+from tracker import constants as C
+from tracker.models import Event, Site, SourceClassification, VisitSession, Visitor
+from tracker.services.attribution import classify
+
+
+def _first_touch_signals(visitor):
+    """最早一条 page_view 的原始信号优先；旧数据无事件属性时回落访客档案字段。"""
+    ev = (Event.objects.filter(visitor=visitor, event_name="page_view")
+          .order_by("occurred_at", "pk").first())
+    props = (ev.properties or {}) if ev else {}
+    referrer = props.get("referrer") if isinstance(props.get("referrer"), str) else ""
+    domain = props.get("referrer_domain") or visitor.first_referrer_domain
+    if not referrer and domain:
+        referrer = f"https://{str(domain).strip().strip('/')}"
+    utm = props.get("utm") if isinstance(props.get("utm"), dict) else None
+    return referrer, (utm or visitor.first_utm or {}), visitor.first_landing_path
+
+
+def _upsert(site, visitor, session, scope, result):
+    vals = {"channel": (result["channel"] or "")[:32],
+            "platform": (result["platform"] or "")[:64],
+            "variant": (result["variant"] or "")[:64],
+            "search_term": (result["search_term"] or "")[:200],
+            "evidence_kind": (result["evidence_kind"] or "")[:32],
+            "resolution": (result["resolution"] or "")[:16],
+            "detail": result["detail"] or {}, "session": session,
+            "rule_version": C.ATTRIBUTION_RULE_VERSION}   # 重算即升到当前规则版本
+    if session is None:
+        SourceClassification.objects.update_or_create(
+            site=site, visitor=visitor, scope=scope, defaults=vals)
+    else:
+        SourceClassification.objects.update_or_create(
+            site=site, session=session, scope=scope,
+            defaults={**vals, "visitor": visitor})
+
+
+class Command(BaseCommand):
+    help = "按当前规则版本全量重算来源分类（first_touch + session_entry，幂等）"
+
+    def add_arguments(self, parser):
+        parser.add_argument("--site", default=None,
+                            help="只重算指定 site_id（缺省全部站点）")
+
+    def handle(self, *args, **opts):
+        sites = Site.objects.all()
+        if opts["site"]:
+            sites = sites.filter(site_id=opts["site"])
+            if not sites.exists():
+                self.stdout.write(self.style.WARNING(
+                    f"site not found: {opts['site']}"))
+                return
+        v_total = s_total = 0
+        for site in sites:
+            domains = site.domains or []
+            for visitor in Visitor.objects.filter(site=site).only("pk", "first_referrer_domain",
+                                                                  "first_utm", "first_landing_path"):
+                referrer, utm, landing = _first_touch_signals(visitor)
+                _upsert(site, visitor, None, "first_touch",
+                        classify(referrer_url=referrer, utm=utm,
+                                 landing_path=landing or "",
+                                 internal_domains=domains))
+                v_total += 1
+            for sess in VisitSession.objects.filter(site=site).select_related("visitor"):
+                referrer = f"https://{sess.referrer_domain.strip().strip('/')}" \
+                    if sess.referrer_domain else ""
+                _upsert(site, sess.visitor, sess, "session_entry",
+                        classify(referrer_url=referrer, utm=sess.utm or {},
+                                 landing_path=sess.landing_path or "",
+                                 internal_domains=domains))
+                s_total += 1
+        self.stdout.write(
+            f"recomputed attribution: {v_total} first_touch, {s_total} session_entry "
+            f"(rule version {C.ATTRIBUTION_RULE_VERSION})")

server/tracker/migrations/0004_sourceclassification.py

@@ -0,0 +1,37 @@
+# Generated by Django 5.2.17 on 2026-09-16 08:52
+
+import django.db.models.deletion
+from django.db import migrations, models
+
+
+class Migration(migrations.Migration):
+
+    dependencies = [
+        ('tracker', '0003_rawbatch'),
+    ]
+
+    operations = [
+        migrations.CreateModel(
+            name='SourceClassification',
+            fields=[
+                ('id', models.BigAutoField(auto_created=True, primary_key=True, serialize=False, verbose_name='ID')),
+                ('scope', models.CharField(max_length=16)),
+                ('channel', models.CharField(blank=True, default='', max_length=32)),
+                ('platform', models.CharField(blank=True, default='', max_length=64)),
+                ('variant', models.CharField(blank=True, default='', max_length=64)),
+                ('search_term', models.CharField(blank=True, default='', max_length=200)),
+                ('evidence_kind', models.CharField(blank=True, default='', max_length=32)),
+                ('resolution', models.CharField(blank=True, default='', max_length=16)),
+                ('detail', models.JSONField(blank=True, default=dict)),
+                ('rule_version', models.CharField(default='v1', max_length=16)),
+                ('computed_at', models.DateTimeField(auto_now=True)),
+                ('session', models.ForeignKey(null=True, on_delete=django.db.models.deletion.CASCADE, related_name='source_classifications', to='tracker.visitsession')),
+                ('site', models.ForeignKey(on_delete=django.db.models.deletion.CASCADE, related_name='source_classifications', to='tracker.site')),
+                ('visitor', models.ForeignKey(null=True, on_delete=django.db.models.deletion.CASCADE, related_name='source_classifications', to='tracker.visitor')),
+            ],
+            options={
+                'db_table': 'source_classification',
+                'indexes': [models.Index(fields=['site', 'visitor', 'scope'], name='source_clas_site_id_58ad9c_idx'), models.Index(fields=['site', 'channel'], name='source_clas_site_id_f08b23_idx')],
+            },
+        ),
+    ]

server/tracker/models.py

@@ -1,6 +1,8 @@
 from django.db import models
 import secrets
 
+from tracker import constants as C
+
 
 def short_public_id() -> str:
     return "site_" + secrets.token_urlsafe(12)[:16]
@@ -131,6 +133,35 @@ class RawBatch(models.Model):
                    models.Index(fields=["site", "-received_at"])]
 
 
+class SourceClassification(models.Model):
+    """来源分类结果（attribution 规则引擎的派生存储）。
+
+    每访客两行 scope：first_touch（首访来源，session=None）/ session_entry（每次新会话入口）。
+    写入两路：ingest 增量（新访客/新会话时 best-effort）+ recompute_attribution 全量重算
+    （幂等 upsert）。resolution 表达证据可信度：resolved/combined/conflict/label_only/unknown——
+    观察到的 referrer 与 UTM 自述矛盾时并列存 detail，不武断取一边。
+    rule_version 关联 constants.ATTRIBUTION_RULE_VERSION：规则升级后可识别存量行需重算。
+    """
+    site = models.ForeignKey(Site, on_delete=models.CASCADE, related_name="source_classifications")
+    visitor = models.ForeignKey(Visitor, on_delete=models.CASCADE, related_name="source_classifications", null=True)
+    session = models.ForeignKey(VisitSession, on_delete=models.CASCADE, related_name="source_classifications", null=True)
+    scope = models.CharField(max_length=16)             # first_touch / session_entry
+    channel = models.CharField(max_length=32, blank=True, default="")
+    platform = models.CharField(max_length=64, blank=True, default="")
+    variant = models.CharField(max_length=64, blank=True, default="")
+    search_term = models.CharField(max_length=200, blank=True, default="")
+    evidence_kind = models.CharField(max_length=32, blank=True, default="")
+    resolution = models.CharField(max_length=16, blank=True, default="")
+    detail = models.JSONField(default=dict, blank=True)
+    rule_version = models.CharField(max_length=16, default=C.ATTRIBUTION_RULE_VERSION)
+    computed_at = models.DateTimeField(auto_now=True)
+
+    class Meta:
+        db_table = "source_classification"
+        indexes = [models.Index(fields=["site", "visitor", "scope"]),
+                   models.Index(fields=["site", "channel"])]
+
+
 class DailyStat(models.Model):
     site = models.ForeignKey(Site, on_delete=models.CASCADE, related_name="daily_stats", db_index=False)   # uq_site_date 前缀覆盖
     date = models.DateField()

server/tracker/services/attribution.py

@@ -0,0 +1,180 @@
+"""来源归因规则引擎：裸域名 → 渠道/平台/搜索词/证据可信度分层识别。
+
+纯函数、无 IO、无 DB——classify 是唯一入口，ingest 增量计算与
+recompute_attribution 历史重算共用同一套规则（改规则必须 bump
+constants.ATTRIBUTION_RULE_VERSION 并重跑重算）。
+
+诚实状态语义（resolution）：
+  resolved    观察到 referrer/点击参数/UA 等硬证据
+  combined    referrer 观察与 UTM 标签互相印证
+  conflict    referrer 结论与 UTM 标签矛盾——观察证据优先，标签并列存 detail
+  label_only  仅 UTM 自述，无观察证据
+  unknown     全空（直接访问）——不猜测手输/书签
+"""
+import re
+from urllib.parse import parse_qsl, urlsplit
+
+from tracker import constants as C
+
+_CTRL_CHARS = re.compile(r"[\x00-\x1f\x7f]")
+_GENERIC_TLDS = ("com",)   # 引擎 host 落在这些 TLD 时不作为国家 variant 显示
+
+
+def _host(url: str) -> str:
+    """提取 referrer host；异常/空 → ""。hostname 自动去端口并小写。"""
+    try:
+        return (urlsplit(url).hostname or "").lower()
+    except ValueError:
+        return ""
+
+
+def _canonical(host: str) -> str:
+    """去掉单个 www. 前缀后做域名匹配（www.bing.com → bing.com）。"""
+    return host[4:] if host.startswith("www.") else host
+
+
+def _domain_match(host: str, key: str) -> bool:
+    """host 是否命中规则键。key 以 "." 结尾（google./yandex.）→ 前缀匹配任意国家站
+    TLD；否则精确或子域（m.facebook.com 命中 facebook.com）。"""
+    h = _canonical(host)
+    k = key.lower()
+    if k.endswith("."):
+        return h == k[:-1] or h.startswith(k)
+    return h == k or h.endswith("." + k)
+
+
+def _tld_variant(host: str) -> str:
+    """引擎 host 的 TLD 国家（www.google.co.ru → "ru"）；通用 TLD 不显示。"""
+    tld = _canonical(host).rsplit(".", 1)[-1]
+    return "" if tld in _GENERIC_TLDS else tld
+
+
+def _clean_term(raw: str) -> str:
+    """搜索词清洗：strip → 剔除控制字符 → 截 100。"""
+    return _CTRL_CHARS.sub("", (raw or "").strip())[:100]
+
+
+def _search_term(engine: str, referrer_url: str) -> str:
+    """按引擎参数名白名单从 referrer URL 提取搜索词。Google 不在表内 → 恒空（绝不虚构）。"""
+    for param in C.SEARCH_TERM_PARAMS.get(engine, []):
+        for k, v in parse_qsl(urlsplit(referrer_url).query, keep_blank_values=True):
+            if k == param and v.strip():
+                return _clean_term(v)
+    return ""
+
+
+def _paid_param(landing_path: str, referrer_url: str) -> tuple[str, str] | None:
+    """付费优先于自然：落地页（或 referrer）query 命中点击参数 → (参数名, 平台名)。"""
+    for source in (landing_path, referrer_url):
+        if not source:
+            continue
+        try:
+            params = [k for k, _ in parse_qsl(urlsplit(source).query,
+                                              keep_blank_values=True)]
+        except ValueError:
+            continue
+        for k in params:
+            if k in C.PAID_CLICK_PARAMS:
+                return k, C.PAID_CLICK_PARAMS[k]
+    return None
+
+
+def _in_app_name(ua: str) -> str:
+    """UA 命中应用内浏览器标记（大小写不敏感）→ 应用显示名；未命中 → ""。"""
+    low = (ua or "").lower()
+    for marker, name in C.INAPP_UA_MARKERS.items():
+        if marker.lower() in low:
+            return name
+    return ""
+
+
+def _utm_channel(label: str) -> str:
+    """UTM 标签尽力归类：已知 AI/引擎/社交标签（裸标签 "baidu"/显示名 "Google" 皆可）
+    映射到对应渠道，未知 → referral（有名来源，可信度由 label_only 单独表达）。"""
+    h = (label or "").lower().strip()
+    if not h:
+        return "referral"
+    for table, channel in ((C.AI_REFERRAL_DOMAINS, "ai_referral"),
+                           (C.SEARCH_ENGINE_DOMAINS, "natural_search"),
+                           (C.SOCIAL_DOMAINS, "social")):
+        for key, name in table.items():
+            if _domain_match(h, key) or key.startswith(h + ".") or h == name.lower():
+                return channel
+    return "referral"
+
+
+def classify(*, referrer_url: str = "", utm: dict | None = None,
+             landing_path: str = "", ua: str = "", internal_domains: list | None = None) -> dict:
+    """把一次到访的原始信号分类为 {channel, platform, variant, search_term,
+    evidence_kind, resolution, detail}。各字段默认空串/空 dict，纯函数无副作用。
+
+    规则序：本站域名去信号 → 付费点击参数 → AI → 搜索引擎 → 社交 → 泛 referral
+    → UTM 证据叠加（combined/conflict/label_only）→ 应用内 UA 叠加/兜底 → direct。
+    """
+    out = {"channel": "", "platform": "", "variant": "", "search_term": "",
+           "evidence_kind": "", "resolution": "", "detail": {}}
+    utm = utm or {}
+    utm_source = str(utm.get("utm_source") or "").strip()
+
+    host = _host(referrer_url)
+    if host and any(_domain_match(host, d) for d in (internal_domains or []) if d):
+        host = ""   # 本站跳转不是外部来源信号
+
+    paid = _paid_param(landing_path, referrer_url)
+    if paid:
+        out.update(channel="paid_search", platform=paid[1],
+                   evidence_kind="observed_referrer", resolution="resolved",
+                   detail={"click_param": paid[0]})
+    elif host:
+        for key, name in C.AI_REFERRAL_DOMAINS.items():
+            if _domain_match(host, key):
+                out.update(channel="ai_referral", platform=name)
+                break
+        else:
+            for key, name in C.SEARCH_ENGINE_DOMAINS.items():
+                if _domain_match(host, key):
+                    out.update(channel="natural_search", platform=name,
+                               variant=_tld_variant(host),
+                               search_term=_search_term(name, referrer_url))
+                    break
+            else:
+                for key, name in C.SOCIAL_DOMAINS.items():
+                    if _domain_match(host, key):
+                        out.update(channel="social", platform=name)
+                        break
+                else:
+                    out.update(channel="referral", platform=host)
+        if not out["resolution"]:
+            out.update(evidence_kind="observed_referrer", resolution="resolved")
+
+    # UTM 证据叠加：观察结论之上只做印证/并列，不覆盖
+    if utm_source:
+        if out["resolution"] == "resolved":
+            if utm_source.lower() == out["platform"].lower():
+                out["evidence_kind"] = "combined"
+            else:   # 矛盾：观察证据保留在顶层，两路证据各自完整并列进 detail
+                out["resolution"] = "conflict"
+                out["detail"] = {**out["detail"],
+                                 "referrer_evidence": {
+                                     "channel": out["channel"], "platform": out["platform"],
+                                     "variant": out["variant"], "search_term": out["search_term"],
+                                     "evidence_kind": "observed_referrer", "referrer_host": host},
+                                 "utm_evidence": {
+                                     "channel": _utm_channel(utm_source), "platform": utm_source,
+                                     "evidence_kind": "utm_label"}}
+        else:   # 无 referrer 观察 → 仅标签：按 utm_source 归类显示，状态诚实标注
+            out.update(channel=_utm_channel(utm_source), platform=utm_source,
+                       evidence_kind="utm_label", resolution="label_only")
+
+    # 应用内浏览器 UA：有来源 → 作为 surface 叠加；无任何来源 → 本身就是来源
+    app = _in_app_name(ua)
+    if app:
+        if out["resolution"]:
+            out["detail"] = {**out["detail"], "in_app": app}
+        else:
+            out.update(channel="in_app", platform=app,
+                       evidence_kind="observed_ua", resolution="resolved")
+
+    if not out["channel"]:
+        out.update(channel="direct", resolution="unknown")   # 不猜测手输/书签
+    return out

server/tracker/services/ingest.py

@@ -21,11 +21,14 @@ from django.db.models.functions import Greatest
 from django.utils import timezone
 
 from tracker import constants as C
-from tracker.models import Visitor, VisitSession, Event
+from tracker.models import SourceClassification, Visitor, VisitSession, Event
 from tracker.services.validation import validate_event
+from tracker.services.attribution import classify
 from tracker.services.geo import resolver as _geo   # 模块级单例（勿实例化，11MB/份）
 
 _FIELD_LEN = {"device_type": 16, "os": 64, "browser": 64, "language": 32, "screen": 32}
+_CLASS_FIELDS = ("channel", "platform", "variant", "search_term",
+                 "evidence_kind", "resolution", "detail")
 
 
 def _clean_utm(d):
@@ -40,6 +43,52 @@ def _geo_field(g, key):
     return v if v not in ("", "0") else ""
 
 
+def _first_pageview_referrer(events) -> str:
+    """本批第一条带 referrer 的 page_view 的完整来源 URL（旧版 SDK 无此字段 → ""）。"""
+    for ev in events:
+        if isinstance(ev, dict) and ev.get("event_name") == "page_view":
+            r = (ev.get("properties") or {}).get("referrer")
+            if isinstance(r, str) and r:
+                return r
+    return ""
+
+
+def _referrer_url(domain, full: str) -> str:
+    """分类入参统一为完整 URL：优先 page_view 的完整 referrer（含搜索词），
+    缺失时用 ctx 的 referrer_domain 合成（丢查询串，搜索词无法找回——诚实降级）。"""
+    if full:
+        return full
+    d = str(domain or "").strip().lower().strip("/")
+    return f"https://{d}" if d else ""
+
+
+def _classify_and_store(*, site, visitor, session, scope, referrer_url,
+                        utm, landing_path, ua):
+    """来源分类落库（best-effort）：任何失败静默吞掉，绝不影响采集主流程。
+    savepoint 隔离——DB 异常不至于标记外层采集事务需回滚。"""
+    try:
+        with transaction.atomic():
+            r = classify(referrer_url=referrer_url, utm=utm or {},
+                         landing_path=landing_path or "", ua=ua or "",
+                         internal_domains=site.domains or [])
+            vals = {"channel": (r["channel"] or "")[:32],
+                    "platform": (r["platform"] or "")[:64],
+                    "variant": (r["variant"] or "")[:64],
+                    "search_term": (r["search_term"] or "")[:200],
+                    "evidence_kind": (r["evidence_kind"] or "")[:32],
+                    "resolution": (r["resolution"] or "")[:16],
+                    "detail": r["detail"] or {}, "session": session}
+            if session is None:   # first_touch：键 site+visitor+scope（每访客一行）
+                SourceClassification.objects.update_or_create(
+                    site=site, visitor=visitor, scope=scope, defaults=vals)
+            else:                 # session_entry：键 site+session+scope（每会话一行）
+                SourceClassification.objects.update_or_create(
+                    site=site, session=session, scope=scope,
+                    defaults={**vals, "visitor": visitor})
+    except Exception:
+        pass
+
+
 def _touch_visitor(site, vid, ctx, ip):
     env = {k: str(ctx.get(k, "") or "")[:n] for k, n in _FIELD_LEN.items()}
     # Task16 联调发现：键名映射缺口——SDK context 发 "timezone"，模型字段是 timezone_name，
@@ -64,10 +113,10 @@ def _touch_visitor(site, vid, ctx, ip):
                        province=_geo_field(g, "province")[:64],
                        city=_geo_field(g, "city")[:64])
     Visitor.objects.filter(pk=v.pk).update(**upd)
-    return v   # 下游只用 pk（FK/子查询），无需 refresh_from_db 多耗一条 SELECT
+    return v, created   # created 供首触来源分类（ingest_batch 消费）
 
 
-def _get_session(site, visitor, sid, ctx, occurred_at):
+def _get_session(site, visitor, sid, ctx, occurred_at, ua="", referrer_url=""):
     """会话权威在 SDK；服务端兜底：同 session_id 族（s1、s1#2…）距最后事件>30min 切分后缀段。
 
     空闲比较必须取整个会话族的最大 ended_at：若只查根段（session_id=sid 精确匹配），
@@ -96,6 +145,13 @@ def _get_session(site, visitor, sid, ctx, occurred_at):
             if not sess:
                 raise
         Visitor.objects.filter(pk=visitor.pk).update(visit_count=F("visit_count") + 1)
+        # 新会话入口来源分类（best-effort）：创建成功与并发撞键重查两条路径都会走到，
+        # 同键 upsert 幂等——赢家请求已写时这只是覆写同值。
+        _classify_and_store(site=site, visitor=visitor, session=sess,
+                            scope="session_entry",
+                            referrer_url=_referrer_url(ctx.get("referrer_domain"), referrer_url),
+                            utm=_clean_utm(ctx.get("utm")),
+                            landing_path=str(ctx.get("path") or ""), ua=ua)
     return sess
 
 
@@ -111,11 +167,19 @@ def ingest_batch(site, payload: dict, ip: str | None = None, ua: str = "") -> di
     if not events or len(events) > C.MAX_EVENTS_PER_BATCH:
         return {"accepted": 0, "duplicated": 0,
                 "rejected": [{"reason": "batch_size_invalid"}]}
-    visitor = _touch_visitor(site, payload["visitor_id"], ctx, ip)
+    visitor, created = _touch_visitor(site, payload["visitor_id"], ctx, ip)
     geo = _geo.resolve(ip) if ip else None
     is_bot = any(k in (ua or "").lower() for k in C.BOT_UA_KEYWORDS)
     if is_bot:
         Visitor.objects.filter(pk=visitor.pk).update(is_bot=True)
+    # 首触来源分类（best-effort）：首批 page_view 的完整 referrer 优先（含搜索词），
+    # 旧 SDK 只有 ctx.referrer_domain 时合成降级。仅在访客首次建档时写一次。
+    pv_referrer = _first_pageview_referrer(events)
+    if created:
+        _classify_and_store(site=site, visitor=visitor, session=None, scope="first_touch",
+                            referrer_url=_referrer_url(ctx.get("referrer_domain"), pv_referrer),
+                            utm=_clean_utm(ctx.get("utm")),
+                            landing_path=str(ctx.get("path") or ""), ua=ua or "")
     ready, session_cache, seen_eids = [], {}, set()
     for i, ev in enumerate(events):
         reason = validate_event(ev)
@@ -131,7 +195,8 @@ def ingest_batch(site, payload: dict, ip: str | None = None, ua: str = "") -> di
         sid = payload["session_id"]
         sess = session_cache.get(sid)
         if not sess:
-            sess = _get_session(site, visitor, sid, ctx, occ)
+            sess = _get_session(site, visitor, sid, ctx, occ, ua=ua or "",
+                                referrer_url=pv_referrer)
             session_cache[sid] = sess      # 按原始 sid 存：切分后 sess.session_id 是 "s1#2"，
                                            # 若按它存，同批后续事件按 "s1" 查会 miss → 每条再切新会话
         ready.append(Event(

server/tracker/services/validation.py

@@ -113,6 +113,10 @@ def validate_event(ev) -> str | None:
         ev["path"] = sanitize_url(ev["path"])
     if len(ev.get("path") or "") > C.MAX_PATH_LEN:
         ev["path"] = ev["path"][:C.MAX_PATH_LEN]
+    # referrer 是完整 URL：同样过参数清洗（email/token 等整键剔除）再截 MAX_STR_LEN。
+    # 放在 str_too_long 检查之后——超长 referrer 仍按 properties_too_long 拒，不静默截半。
+    if ev["event_name"] == "page_view" and isinstance(props.get("referrer"), str) and props["referrer"]:
+        props["referrer"] = sanitize_url(props["referrer"])[:C.MAX_STR_LEN]
     if len(ev.get("title") or "") > C.MAX_TITLE_LEN:
         ev["title"] = ev["title"][:C.MAX_TITLE_LEN]
     return None

server/tracker/tests/test_attribution.py

@@ -0,0 +1,183 @@
+"""来源归因规则引擎（tracker/services/attribution.classify）。
+
+纯函数用例（无 DB）+ 重算命令/入库集成为带 db 标记的独立用例。
+诚实状态语义：resolved=观察到 referrer 证据；label_only=仅 UTM 自述；
+conflict=两证据矛盾并列；unknown=直接访问不猜测。
+"""
+import pytest
+from django.core.management import call_command
+from django.utils import timezone
+
+from tracker.models import Site, SourceClassification, VisitSession, Visitor
+from tracker.services.attribution import classify
+
+NOW = timezone.now()
+
+
+# ---------------------------------------------------------------- 引擎纯函数
+def test_bing_search_term_extracted():
+    r = classify(referrer_url="https://www.bing.com/search?q=%E6%8C%96%E6%8E%98%E6%9C%BA&m=1",
+                 landing_path="/pricing")
+    assert r["channel"] == "natural_search"
+    assert r["platform"] == "Bing"
+    assert r["search_term"] == "挖掘机"
+    assert r["evidence_kind"] == "observed_referrer"
+    assert r["resolution"] == "resolved"
+
+
+def test_google_never_fabricates_search_term():
+    r = classify(referrer_url="https://www.google.com/search?q=cats")
+    assert r["channel"] == "natural_search" and r["platform"] == "Google"
+    assert r["search_term"] == ""            # Google 不在 SEARCH_TERM_PARAMS 表 = 恒空，绝不虚构
+    assert r["variant"] == ""                # google.com → TLD "com" 不显示
+
+
+def test_google_country_tld_as_variant():
+    r = classify(referrer_url="https://www.google.co.ru/search?q=x")
+    assert r["platform"] == "Google" and r["variant"] == "ru"
+
+
+def test_baidu_wd_chinese_term_and_word_fallback():
+    r = classify(referrer_url="https://www.baidu.com/s?wd=挖掘机价格")
+    assert r["platform"] == "百度" and r["search_term"] == "挖掘机价格"
+    r2 = classify(referrer_url="https://m.baidu.com/s?word=租赁")   # 子域 + word 参数兜底
+    assert r2["platform"] == "百度" and r2["search_term"] == "租赁"
+
+
+def test_chatgpt_referral():
+    r = classify(referrer_url="https://chatgpt.com/c/abc")
+    assert r["channel"] == "ai_referral" and r["platform"] == "ChatGPT"
+    assert r["resolution"] == "resolved"
+    r2 = classify(referrer_url="https://chat.chatgpt.com/")   # 子域同样命中
+    assert r2["channel"] == "ai_referral" and r2["platform"] == "ChatGPT"
+
+
+def test_social_twitter_short_link():
+    r = classify(referrer_url="https://t.co/AbCdEf")
+    assert r["channel"] == "social" and r["platform"] == "Twitter/X"
+
+
+def test_generic_referral_keeps_host_as_platform():
+    r = classify(referrer_url="https://blog.example.com/post/1")
+    assert r["channel"] == "referral" and r["platform"] == "blog.example.com"
+    assert r["evidence_kind"] == "observed_referrer" and r["resolution"] == "resolved"
+
+
+def test_paid_click_param_wins_over_natural_search():
+    r = classify(referrer_url="https://www.bing.com/search?q=x",
+                 landing_path="/offer?gclid=EAIaIQ")
+    assert r["channel"] == "paid_search" and r["platform"] == "Google Ads"   # 付费优先于自然
+    assert r["detail"].get("click_param") == "gclid"
+    r2 = classify(landing_path="/?fbclid=IwAR0")                            # 仅落地页带点击参数
+    assert r2["channel"] == "paid_search" and r2["platform"] == "Meta Ads"
+
+
+def test_paid_param_in_referrer_url():
+    r = classify(referrer_url="https://example.com/ad?msclkid=xyz", landing_path="/")
+    assert r["channel"] == "paid_search" and r["platform"] == "Bing Ads"
+
+
+def test_micromessenger_ua_without_referrer_is_in_app():
+    r = classify(ua="Mozilla/5.0 (Linux; Android 14) MicroMessenger/8.0.47")
+    assert r["channel"] == "in_app" and r["platform"] == "微信"
+    assert r["resolution"] == "resolved"
+
+
+def test_in_app_marker_overlays_existing_source_in_detail():
+    r = classify(referrer_url="https://www.bing.com/search?q=x",
+                 ua="Mozilla/5.0 ... MicroMessenger/8.0.47")
+    assert r["channel"] == "natural_search"          # 来源结论不被 UA 覆盖
+    assert r["detail"].get("in_app") == "微信"         # 应用内浏览器作为 surface 叠加进 detail
+
+
+def test_internal_referrer_treated_as_no_signal():
+    r = classify(referrer_url="https://www.a.com/other-page", internal_domains=["a.com"])
+    assert r["channel"] == "direct" and r["resolution"] == "unknown"
+    r2 = classify(referrer_url="https://sub.a.com/x", utm={"utm_source": "newsletter"},
+                  internal_domains=["a.com"])        # 本站跳转后仍带 UTM → 仅剩标签证据
+    assert r2["resolution"] == "label_only" and r2["platform"] == "newsletter"
+
+
+def test_utm_matching_referrer_upgrades_to_combined():
+    r = classify(referrer_url="https://www.bing.com/search?q=x", utm={"utm_source": "bing"})
+    assert r["channel"] == "natural_search" and r["evidence_kind"] == "combined"
+
+
+def test_utm_contradicting_referrer_is_conflict_with_both_evidences():
+    r = classify(referrer_url="https://facebook.com/l/abc", utm={"utm_source": "google"})
+    assert r["channel"] == "social" and r["platform"] == "Facebook"   # 观察证据优先于自述标签
+    assert r["resolution"] == "conflict"
+    assert r["evidence_kind"] == "observed_referrer"
+    ref_ev, utm_ev = r["detail"]["referrer_evidence"], r["detail"]["utm_evidence"]
+    assert ref_ev["platform"] == "Facebook" and ref_ev["channel"] == "social"
+    assert utm_ev["platform"] == "google" and utm_ev["evidence_kind"] == "utm_label"
+
+
+def test_utm_only_is_label_only():
+    r = classify(utm={"utm_source": "newsletter"})
+    assert r["channel"] == "referral"               # 未知标签 → 有名 referral，但状态诚实标注
+    assert r["platform"] == "newsletter"
+    assert r["resolution"] == "label_only" and r["evidence_kind"] == "utm_label"
+    r2 = classify(utm={"utm_source": "baidu"})      # 已知引擎标签 → 归入对应渠道（仍 label_only）
+    assert r2["channel"] == "natural_search" and r2["platform"] == "baidu"
+    assert r2["resolution"] == "label_only"
+
+
+def test_empty_everything_is_direct_unknown():
+    r = classify()
+    assert r == {"channel": "direct", "platform": "", "variant": "",
+                 "search_term": "", "evidence_kind": "", "resolution": "unknown",
+                 "detail": {}}
+
+
+def test_search_term_cleaned_and_truncated():
+    long = "词" * 150
+    r = classify(referrer_url=f"https://www.bing.com/search?q=%20%20{long}%20")
+    assert r["search_term"] == "词" * 100           # strip + 截 100
+    r2 = classify(referrer_url="https://www.bing.com/search?q=a%00b%01c")  # 控制字符剔除
+    assert r2["search_term"] == "abc"
+
+
+def test_garbage_referrer_url_does_not_crash():
+    r = classify(referrer_url="http://[invalid", landing_path="/x")
+    assert r["channel"] in ("referral", "direct")   # 不崩，落到兜底分支即可
+
+
+# ------------------------------------------------------- 入库 + 重算集成
+@pytest.mark.django_db
+def test_ingest_writes_first_touch_and_session_entry():
+    from tracker.services.ingest import ingest_batch
+    site = Site.objects.create(name="s", domains=["a.com"])
+    payload = {"visitor_id": "v1", "session_id": "s1", "context": {
+        "referrer_domain": "www.bing.com", "utm": {}, "path": "/landing"},
+        "events": [{"event_id": "e1", "event_name": "page_view", "occurred_at": NOW,
+                    "path": "/landing", "title": "T",
+                    "properties": {"referrer": "https://www.bing.com/search?q=%E6%8C%96%E6%8E%98%E6%9C%BA"}}]}
+    ingest_batch(site, payload, ip="114.114.114.114", ua="Mozilla/5.0")
+    ft = SourceClassification.objects.get(site=site, visitor__visitor_id="v1", scope="first_touch")
+    assert ft.session is None
+    assert ft.channel == "natural_search" and ft.platform == "Bing"
+    assert ft.search_term == "挖掘机" and ft.rule_version == "v1"
+    se = SourceClassification.objects.get(site=site, scope="session_entry")
+    assert se.session.session_id == "s1"
+    assert se.channel == "natural_search" and se.platform == "Bing"
+
+
+@pytest.mark.django_db
+def test_recompute_attribution_idempotent():
+    from tracker.services.ingest import ingest_batch
+    site = Site.objects.create(name="s", domains=["a.com"])
+    payload = {"visitor_id": "v1", "session_id": "s1", "context": {
+        "referrer_domain": "www.bing.com", "utm": {}, "path": "/landing"},
+        "events": [{"event_id": "e1", "event_name": "page_view", "occurred_at": NOW,
+                    "path": "/landing", "title": "T",
+                    "properties": {"referrer": "https://www.bing.com/search?q=ok"}}]}
+    ingest_batch(site, payload, ip="114.114.114.114", ua="Mozilla/5.0")
+    SourceClassification.objects.all().delete()      # 清掉增量写入，验证重算可全量重建
+    call_command("recompute_attribution", stdout=None)
+    n1 = SourceClassification.objects.count()
+    assert n1 == 2                                   # first_touch + session_entry 各一
+    ft = SourceClassification.objects.get(scope="first_touch")
+    assert ft.channel == "natural_search" and ft.search_term == "ok"
+    call_command("recompute_attribution", stdout=None)   # 第二遍：upsert 不重复不漂移
+    assert SourceClassification.objects.count() == n1

server/tracker/tests/test_dashboard.py

@@ -3,7 +3,7 @@ from datetime import timedelta
 from django.contrib.auth import get_user_model
 from django.utils import timezone
 from rest_framework.test import APIClient
-from tracker.models import Event, Site, VisitSession, Visitor
+from tracker.models import Event, Site, SourceClassification, VisitSession, Visitor
 from tracker.services.stats import refresh_daily
 
 pytestmark = pytest.mark.django_db
@@ -104,6 +104,33 @@ def test_overview_trend_uses_dailystat_history(client, member_user):
     assert d["trend"][-1]["pv"] == 0   # 今天无事件，实时覆盖为 0
 
 
+def test_overview_channels_aggregation(client, member_user):
+    """channels：当天会话按 session_entry 分类的 channel 聚合（与 sources 并存，互不影响）。"""
+    _auth(client, member_user)
+    site = Site.objects.create(name="s", domains=["a.com"])
+    v1 = _visitor(site, "v1", last_seen=timezone.now())
+    v2 = _visitor(site, "v2", last_seen=timezone.now())
+    s1 = _session(site, v1, "s1", referrer="www.bing.com")
+    s2 = _session(site, v1, "s2", referrer="www.bing.com")
+    s3 = _session(site, v2, "s3", referrer="chatgpt.com")
+    s4 = _session(site, v2, "s4", referrer="")                      # 直接访问
+    SourceClassification.objects.create(site=site, visitor=v1, session=s1,
+        scope="session_entry", channel="natural_search", platform="Bing")
+    SourceClassification.objects.create(site=site, visitor=v1, session=s2,
+        scope="session_entry", channel="natural_search", platform="Bing")
+    SourceClassification.objects.create(site=site, visitor=v2, session=s3,
+        scope="session_entry", channel="ai_referral", platform="ChatGPT")
+    SourceClassification.objects.create(site=site, visitor=v2, session=s4,
+        scope="session_entry", channel="direct", resolution="unknown")
+    _events(site, v1, s1, 2, "/", prefix="a")
+
+    d = client.get("/api/admin/dashboard/overview").json()["data"]
+    assert d["channels"] == [{"channel": "natural_search", "sessions": 2},
+                             {"channel": "ai_referral", "sessions": 1},
+                             {"channel": "direct", "sessions": 1}]   # 倒序 top8，平局按名稳定
+    assert d["sources"][0] == {"source": "www.bing.com", "sessions": 2}  # 原口径保留
+
+
 def test_overview_site_filter_and_404(client, member_user):
     _auth(client, member_user)
     site_a = Site.objects.create(name="a", domains=["a.com"])

server/tracker/tests/test_ingest.py

@@ -142,3 +142,23 @@ def test_context_timezone_persisted_to_visitor(db):
     b["context"]["timezone"] = "Asia/Shanghai"
     ingest_batch(site, b)
     assert Visitor.objects.get(visitor_id="v1").timezone_name == "Asia/Shanghai"
+
+# 15. 归因分类 best-effort：引擎抛异常不影响采集入库（first_touch/session_entry 双路）
+def test_attribution_failure_never_breaks_ingest(db, monkeypatch):
+    from tracker.services import ingest as ingest_mod
+    site = Site.objects.create(name="s", domains=["a.com"])
+    monkeypatch.setattr(ingest_mod, "classify", lambda **kw: (_ for _ in ()).throw(RuntimeError("boom")))
+    r = ingest_batch(site, _batch([_ev(1)]))
+    assert r["accepted"] == 1 and Event.objects.count() == 1
+    assert Visitor.objects.count() == 1 and VisitSession.objects.count() == 1
+
+# 16. 首触分类用本批第一条 page_view 的完整 referrer（搜索词可恢复），非 ctx 裸域名
+def test_first_touch_classification_uses_full_pageview_referrer(db):
+    from tracker.models import SourceClassification
+    site = Site.objects.create(name="s", domains=["a.com"])
+    b = _batch([_ev(1, properties={"referrer": "https://m.baidu.com/s?wd=挖掘机"})])
+    b["context"]["utm"] = {}     # 默认 fixture 带 utm_source=google，与百度矛盾 → conflict
+    ingest_batch(site, b)
+    ft = SourceClassification.objects.get(scope="first_touch")
+    assert ft.channel == "natural_search" and ft.platform == "百度"
+    assert ft.search_term == "挖掘机" and ft.resolution == "resolved"

server/tracker/tests/test_validation.py

@@ -37,6 +37,21 @@ def test_sanitize_url_drops_sensitive_params():
     assert "email=" not in u and "token=" not in u and "utm_source=google" in u
     assert "monkey=1" in u          # fullmatch：含 key 子串的正常参数不误伤
 
+
+def test_pageview_referrer_url_sanitized():
+    """page_view 的 properties.referrer 是完整 URL：敏感参数整键剔除后写回。"""
+    ev = _ev(properties={"referrer": "https://www.bing.com/search?q=x&email=a@b.com&token=t"})
+    assert validate_event(ev) is None
+    assert ev["properties"]["referrer"] == "https://www.bing.com/search?q=x"
+
+
+def test_pageview_referrer_truncated_after_reencode():
+    """清洗重编码（中文 → %XX 九倍膨胀）可能超出原长：出门统一截 MAX_STR_LEN=500。"""
+    ev = _ev(properties={"referrer": "https://www.bing.com/search?q=" + "词" * 200})
+    assert validate_event(ev) is None              # 原文 227 字符过长度门
+    ref = ev["properties"]["referrer"]
+    assert len(ref) == 500 and ref.startswith("https://www.bing.com/search?q=%E8%AF%8D")
+
 def test_nested_sensitive_key_rejected():
     assert validate_event(_ev(properties={"utm": {"utm_a": 1, "user_password": "x"}})) == "sensitive_field"  # utm 在白名单内，嵌套敏感键仍须命中
 

server/tracker/tests/test_visitors.py

@@ -3,7 +3,7 @@ from datetime import timedelta
 from django.contrib.auth import get_user_model
 from django.utils import timezone
 from rest_framework.test import APIClient
-from tracker.models import Event, Site, VisitSession, Visitor
+from tracker.models import Event, Site, SourceClassification, VisitSession, Visitor
 
 pytestmark = pytest.mark.django_db
 User = get_user_model()
@@ -77,7 +77,8 @@ def test_list_fields_and_default_bot_excluded(client, member_user):
     assert set(row) == {"visitor_id", "site_id", "site_name", "country", "province", "city",
                         "device_type", "os", "browser", "language", "first_seen_at",
                         "last_seen_at", "visit_count", "active_days", "session_count",
-                        "total_stay_ms", "first_source", "last_source", "online"}
+                        "total_stay_ms", "first_source", "last_source", "online",
+                        "channel", "platform", "variant", "search_term", "resolution"}
     assert row["country"] == "中国" and row["os"] == "Windows" and row["browser"] == "Chrome"
     assert row["session_count"] == 2
     assert row["total_stay_ms"] == 150000            # 120000+30000，有效停留口径
@@ -319,3 +320,63 @@ def test_timeline_orphan_events_without_session(client, member_user):
     sess = r.json()["data"]["days"][0]["sessions"][0]
     assert sess["session_id"] is None and sess["started_at"] is None
     assert [e["event_name"] for e in sess["events"]] == ["sdk_loaded", "page_stay"]
+
+
+def _classify(site, visitor, scope, channel, platform="", session=None, **kw):
+    return SourceClassification.objects.create(
+        site=site, visitor=visitor, session=session, scope=scope,
+        channel=channel, platform=platform, **kw)
+
+
+def test_list_attribution_fields_and_channel_filter(client, member_user):
+    _auth(client, member_user)
+    site = Site.objects.create(name="s", domains=["a.com"])
+    v1 = _visitor(site, "v1", last_seen=timezone.now())
+    v2 = _visitor(site, "v2", last_seen=timezone.now() - timedelta(hours=1))
+    v3 = _visitor(site, "v3", last_seen=timezone.now() - timedelta(hours=2))  # 无分类行
+    _classify(site, v1, "first_touch", "natural_search", "Bing",
+              search_term="挖掘机", evidence_kind="observed_referrer",
+              resolution="resolved")
+    _classify(site, v2, "first_touch", "ai_referral", "ChatGPT",
+              evidence_kind="observed_referrer", resolution="resolved")
+
+    d = client.get("/api/admin/visitors").json()["data"]
+    by_id = {x["visitor_id"]: x for x in d["list"]}
+    assert by_id["v1"]["channel"] == "natural_search"
+    assert by_id["v1"]["platform"] == "Bing" and by_id["v1"]["search_term"] == "挖掘机"
+    assert by_id["v1"]["variant"] == "" and by_id["v1"]["resolution"] == "resolved"
+    assert by_id["v3"]["channel"] == ""            # 无分类行 → 空串（前端回落 first_source 展示）
+
+    d = client.get("/api/admin/visitors", {"channel": "natural_search"}).json()["data"]
+    assert d["total"] == 1 and [x["visitor_id"] for x in d["list"]] == ["v1"]
+    d = client.get("/api/admin/visitors", {"channel": "direct"}).json()["data"]
+    assert d["total"] == 0                         # 无分类访客不落进任何渠道桶
+    d = client.get("/api/admin/visitors", {"channel": "ai_referral"}).json()["data"]
+    assert [x["visitor_id"] for x in d["list"]] == ["v2"]
+
+
+def test_detail_attribution_block(client, member_user):
+    _auth(client, member_user)
+    site = Site.objects.create(name="s", domains=["a.com"])
+    v = _visitor(site, "v1")
+    s1 = _session(site, v, "s1", at=NOW - timedelta(hours=2),
+                  referrer_domain="www.bing.com")
+    s2 = _session(site, v, "s2", at=NOW, referrer_domain="chatgpt.com")
+    _classify(site, v, "first_touch", "natural_search", "Bing",
+              search_term="挖掘机", evidence_kind="observed_referrer",
+              resolution="resolved")
+    _classify(site, v, "session_entry", "natural_search", "Bing",
+              session=s1, evidence_kind="observed_referrer", resolution="resolved")
+    _classify(site, v, "session_entry", "ai_referral", "ChatGPT",
+              session=s2, evidence_kind="observed_referrer", resolution="resolved")
+
+    r = client.get("/api/admin/visitors/v1")
+    assert r.status_code == 200
+    attr = r.json()["data"]["attribution"]
+    ft = attr["first_touch"]
+    assert ft["channel"] == "natural_search" and ft["platform"] == "Bing"
+    assert ft["search_term"] == "挖掘机" and ft["resolution"] == "resolved"
+    assert ft["evidence_kind"] == "observed_referrer" and ft["rule_version"] == "v1"
+    assert ft["computed_at"] and ft["detail"] == {}
+    assert [e["session_id"] for e in attr["session_entries"]] == ["s2", "s1"]  # 最近在前
+    assert attr["session_entries"][0]["channel"] == "ai_referral"