fix(benchmark): fair to reasoning models, fair to router degradation, real sub-cent costs

Three false signals in benchmark results, all seen on a live run:

- Reasoning models were scored ❌ because the per-test max_tokens cap cut
  them mid-<think> before the actual answer. Every run now gets +256 tokens
  of headroom; non-thinking models stop at EOS as before, thinking ones get
  to finish (math-mult on nanbeige: false 0/1 → honest 1/1).
- When the router degraded (timeout → fallback chain) the table showed
  "✗ degraded" and counted it against classification accuracy. There was no
  classification to judge — such rows now show a muted "⚠ router degraded"
  with an explanatory tooltip and are excluded from the accuracy KPI.
- Costs below $0.0001 (cheap flash models) collapsed to "$0.0000". money
  formatting now grows precision until two significant digits survive
  ($0.000035 instead of $0.0000).

2 new strings in all 7 languages.

web/js/ui.js

@@ -100,13 +100,16 @@ export function confirmDialog(message) {
 export function moneyEl(n) {
   if (n == null) return document.createTextNode("—");
   if (n === 0) return document.createTextNode("$0");
-  const s = n.toFixed(4);
+  // sub-0.0001 costs (cheap flash models) would collapse to $0.0000 — grow
+  // precision until two significant digits survive
+  const s = Math.abs(n) >= 0.0001 ? n.toFixed(4) : n.toFixed(Math.min(8, 1 - Math.floor(Math.log10(Math.abs(n)))));
   return h("span", { class: "money" }, "$" + s.slice(0, -2), h("small", { class: "money-tail" }, s.slice(-2)));
 }
 
 export function money(n) {
   if (n == null) return "—";
   if (n === 0) return "$0";
+  if (Math.abs(n) > 0 && Math.abs(n) < 0.0001) return "$" + n.toFixed(Math.min(8, 1 - Math.floor(Math.log10(Math.abs(n)))));
   // up to 4 decimals, trailing zeros trimmed (but keep money-like 2 minimum)
   let s = n.toFixed(4).replace(/0+$/, "").replace(/\.$/, "");
   if (/\.\d$/.test(s)) s += "0";