perf: reuse per-node matches + skip scans on node-less mutations in content.js (closes task/content-scan-efficiency)

Source: task/content-scan-efficiency (plan/steward-linker-quality) — the observer re-scanned on unrelated churn and the regex ran twice per node per scan.

- scan() now caches each text node's allMatchesIn() result in a per-scan
  Map<node, matches>; wrapNode reuses that array instead of re-running the
  regex battery, so the regex runs once per node per scan, not twice.
- The MutationObserver stays broad (documentElement, subtree) but
  short-circuits: it only schedules a scan when a mutation batch actually
  ADDED an element/text node. SPA navigation and "load more" add nodes and
  still scan; removal-only / attribute / characterData churn (vote counts,
  timers, hovercards) no longer fires the 600ms full scan.
- Behavior-preserving: same matches, same wrapping, same injected-link set;
  rsl-link idempotency skip + running/pending serialization latch + debounce
  all intact.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

extension/content.js

@@ -102,9 +102,11 @@
     return a;
   }
 
-  function wrapNode(node) {
+  // `matches` is the node's allMatchesIn(text) result, computed once per scan in scan()
+  // and reused here — the regex battery runs once per node per scan, not twice.
+  function wrapNode(node, matches) {
     const text = node.nodeValue;
-    const ms = Extract.allMatchesIn(text).filter((m) => {
+    const ms = matches.filter((m) => {
       const links = resolved.get(m.name);
       return links && ((links.primary && links.primary.url) || links.spotify);
     });
@@ -164,8 +166,17 @@
     log("scan: root", root.tagName || root.nodeName, "| text nodes", nodes.length);
     if (!nodes.length) return;
 
+    // Run the regex battery ONCE per node this scan: cache each node's full match array in
+    // matchesByNode and derive the candidate `seen` set from it, so wrapNode can reuse the same
+    // array instead of re-running allMatchesIn. Scoped to this scan (rebuilt each call) — node text
+    // can change between scans, so it is never persisted.
+    const matchesByNode = new Map();
     const seen = new Set();
-    for (const node of nodes) for (const m of Extract.allMatchesIn(node.nodeValue)) seen.add(m.name);
+    for (const node of nodes) {
+      const ms = Extract.allMatchesIn(node.nodeValue);
+      matchesByNode.set(node, ms);
+      for (const m of ms) seen.add(m.name);
+    }
     log("candidates found:", seen.size);
 
     const ask = [...seen].filter((c) => !resolved.has(c));
@@ -180,7 +191,7 @@
         if (Object.prototype.hasOwnProperty.call(res, c)) resolved.set(c, res[c] || null);
       }
     }
-    for (const node of nodes) if (node.isConnected) wrapNode(node);
+    for (const node of nodes) if (node.isConnected) wrapNode(node, matchesByNode.get(node));
 
     const total = document.querySelectorAll("a.rsl-link").length;
     api.runtime.sendMessage({ type: "badge", count: total }).catch(() => {});
@@ -208,8 +219,25 @@
 
   const onThread = () => /\/comments\//.test(location.pathname);
 
+  // Only ADDED element/text nodes can introduce new comment text to link, so a mutation batch
+  // that added none (pure attribute/characterData churn on existing nodes — vote counts, timers,
+  // hovercard text) can't change the link set and is skipped. SPA navigation and "load more" both
+  // ADD nodes, so they still trigger a scan. Conservative: any addedNodes of a relevant type
+  // schedules — we never inspect the added node's contents (late-rendered text would be missed).
+  function addedRelevantNodes(records) {
+    for (const r of records) {
+      for (const n of r.addedNodes) {
+        if (n.nodeType === Node.ELEMENT_NODE || n.nodeType === Node.TEXT_NODE) return true;
+      }
+    }
+    return false;
+  }
+
   log("content script loaded:", location.href, "| thread page:", onThread());
   if (onThread()) schedule();
-  new MutationObserver(debounce(() => { if (onThread()) schedule(); }, 600))
+  // The observer stays broad (documentElement, subtree) so a SPA nav that replaces the comment-tree
+  // root is still seen; the short-circuit below cuts the churn-scans without narrowing the root.
+  const scheduleScan = debounce(() => { if (onThread()) schedule(); }, 600);
+  new MutationObserver((records) => { if (addedRelevantNodes(records)) scheduleScan(); })
     .observe(document.documentElement, { childList: true, subtree: true });
 })();