fix(billing): 未定价按天统计，让「已修好」和「还在漏」不再长得一样

计费页只有一个终身计数器 unpriced_reqs，实测 6378 —— 占 18981 请求的 34%。
但这个数字回答不了运维唯一会问的问题：「还在漏吗？」

原因是它把两件事混在一起：加规则之前的历史欠账，和加规则之后仍在发生的漏计。
修好一个源之后终身计数纹丝不动，于是仪表盘上「未定价 6378」和一个仍在
什么都没记的系统完全无法区分。

改动：
- 钩子按 UTC 天写入 unpriced_by_day{day:{reqs, models}}。终身计数保留
  （报表不该改写历史），天桶是新增的观察维度。
- 页面加「未定价（近 7 天）」一行。为 0 时额外显示「近 7 天无」——这时候
  零本身就是结论：加的规则确实生效了。
- 没有天桶数据时返回 null 而不是 0。旧的 state 文件说「未知」，
  不该假装有一份干净记录。

生产实测：全部视图 终身 6378 / 近 7 天 2；日视图隐藏终身计数（null）但保留
近 7 天 —— 周期视图原本把两者都置 null，因为那时根本没有天桶。

失败的请求同样计入（上游 502 的 gpt-5.5 也进了天桶）：一次失败的请求照样
消耗上游配额，只是没拿到 token，单看 token 维度会漏掉。

判据：两个不同日期的请求必须落到两个天桶且终身计数不重复累加；变异成
「总记今天」即变红（只判首日会漏掉这一点）。

Co-Authored-By: ModelRouter <noreply@modelrouter.dev>

internal/lua/plugins/billing.lua

@@ -442,6 +442,21 @@ function plugin.on_request_end(payload)
     s.unpriced_reqs = s.unpriced_reqs + 1
     local m = payload.model or "?"
     s.unpriced_models[m] = (s.unpriced_models[m] or 0) + 1
+
+  -- Day-keyed unpriced tally. The lifetime counters above cannot answer the
+  -- question an operator actually has — "is it STILL happening?" — because they
+  -- keep the debt from before a pricing rule was added. After pricing a source
+  -- that was mispriced, the lifetime number never moves, so "unpriced: 6376" on
+  -- a dashboard is indistinguishable from a system that is still billing
+  -- nothing. Counting by day makes the fix observable: the offending day keeps
+  -- its number, new days start at zero.
+  if s.unpriced_by_day == nil then s.unpriced_by_day = {} end
+  -- dayKey wants epoch SECONDS; payload.time is milliseconds.
+  local uday = dayKey(math.floor((tonumber(payload.time) or (os.time() * 1000)) / 1000))
+  local ubucket = s.unpriced_by_day[uday]
+  if ubucket == nil then ubucket = { reqs = 0, models = {} }; s.unpriced_by_day[uday] = ubucket end
+  ubucket.reqs = ubucket.reqs + 1
+  ubucket.models[m] = (ubucket.models[m] or 0) + 1
   end
   if s.degraded_reqs == nil then s.degraded_reqs = 0 end
   -- Degradation is counted here rather than in the chain_step hook because
@@ -517,7 +532,7 @@ plugin.ui = {
     global = {
       "total", "by_source", "by_model", "by_day", "by_day_src",
       "by_day_model", "by_tier_served", "skip_reasons", "degraded_reqs",
-      "unpriced_models", "unpriced_reqs", "started",
+      "unpriced_models", "unpriced_reqs", "unpriced_by_day", "started",
     },
     depth = { by_key = 1, by_day_key = 2 },
   },
@@ -593,6 +608,7 @@ plugin.ui = {
     en: {
       total: "Total", requests: "Requests", degraded: "Degraded",
       unpriced: "Unpriced", prompt: "Prompt tokens",
+      unpriced7d: "Unpriced (7d)", unpricedClean: "none in last 7 days",
       completion: "Completion tokens", failures: "Failures",
       cacheRate: "Cache hit rate", cacheTokens: "Cache read tokens",
       perSource: "Per source", perModel: "Per model",
@@ -614,6 +630,7 @@ plugin.ui = {
     zh: {
       total: "总开销", requests: "请求数", degraded: "降级",
       unpriced: "未定价", prompt: "输入 tokens",
+      unpriced7d: "未定价（近 7 天）", unpricedClean: "近 7 天无",
       completion: "输出 tokens", failures: "失败",
       cacheRate: "缓存命中率", cacheTokens: "缓存读取 tokens",
       perSource: "按源", perModel: "按模型",
@@ -734,13 +751,35 @@ plugin.ui = {
     });
     return out;
   }
+  // recentUnpriced counts unpriced requests over the last n days. When the
+  // result is zero THAT is the useful signal: every pricing rule added since
+  // then is holding. Null when there is no day data at all (a state file
+  // written before the buckets existed) so the UI can say "unknown" rather than
+  // claim a clean record it cannot support.
+  function recentUnpriced(st, n) {
+    var byDay = st && st.unpriced_by_day;
+    if (!byDay) return null;
+    var cutoff = Date.now() - n * 86400000;
+    var sum = 0, seen = false;
+    Object.keys(byDay).forEach(function (d) {
+      var t = Date.parse(d + "T00:00:00Z");
+      if (isNaN(t)) return;
+      seen = true;
+      if (t >= cutoff) sum += (byDay[d].reqs || 0);
+    });
+    return seen ? sum : null;
+  }
+
   // rescale returns the state as the selected period presents it. Degraded and
   // unpriced counters are lifetime-only (there is no day-keyed tally for them)
   // so they are returned as null and the UI hides them instead of printing a
   // lifetime number beside a per-window cost — the exact mismatch the period
   // view exists to remove.
   function rescale(st) {
-    if (PERIOD === "all" || !st.by_day) return st;
+    if (PERIOD === "all" || !st.by_day) {
+      if (st) st.unpriced_recent = recentUnpriced(st, 7);
+      return st;
+    }
     var days = daysInWindow(PERIOD, utcToday());
     var total = zeroBucket();
     var byDay = {};
@@ -758,6 +797,10 @@ plugin.ui = {
       by_day: byDay,
       degraded_reqs: null,
       unpriced_reqs: null,
+      // The lifetime unpriced total is null here for the same reason, but the
+      // day buckets are not — they are what lets the card report "0 in the last
+      // 7 days" instead of leaving a settled debt indistinguishable from a leak.
+      unpriced_recent: recentUnpriced(st, 7),
     };
     return out;
   }
@@ -838,6 +881,15 @@ plugin.ui = {
       // print a lifetime number next to a per-window cost.
       view.degraded_reqs != null ? [T.degraded, fmtInt(view.degraded_reqs)] : null,
       view.unpriced_reqs != null ? [T.unpriced, fmtInt(view.unpriced_reqs)] : null,
+      // "Unpriced (7d)" answers what the lifetime number cannot: is it STILL
+      // happening? A settled debt and a live leak print the same lifetime total.
+      view.unpriced_recent != null
+        ? [
+            T.unpriced7d,
+            fmtInt(view.unpriced_recent),
+            view.unpriced_recent === 0 ? T.unpricedClean : null,
+          ]
+        : null,
       [T.prompt, fmtInt(t.prompt_tokens || 0)],
       [T.completion, fmtInt(t.completion_tokens || 0)],
       [T.failures, t.failures || 0],