fix(search): nobody types the sigil, so lastSeg should not require it

`lte` ranked 27th for `FilterInput.$lte` behind 26 titles that merely contain
those three letters inside aLTErnative and fiLTEr: /compare/ at 700,
FilteredChannels 565, FILTER_OPS 550, against the target's 438. The same shape as
"rate" leading with generate_client.

The cause was not the matcher. W.lastSeg exists for exactly this query — "send"
for RedisQueue.send, 700 points, the strongest applicable signal — but the last
segment of `filterinput.$lte` folds to `$lte`, which never equals `lte`. The
record fell through to W.substring at 320 and lost to containment. The sigil is
TypeScript's, not the reader's. Comparing sigil-insensitively covers `$lte`,
`$gte`, `$in` and `_default`: 77 of 1,152 API records have a sigil-prefixed last
segment.

Two grading attempts came first and are recorded because both looked right and
neither worked. Grading scanFor so `\blte\b` outscores `[a-z]lte[a-z]` moved it
only 27 -> 14 and cost natural macro 0.1. Grading W.substring the same way did
nothing at all: `direct` is a Math.max over the element score, and for a long
title the element score already exceeds 320, so the substring weight was never
what decided it. The lever was a signal that was missing, not one that was
mis-weighted.

                        master     after
  recall@6 (agents)      97.2%     99.5%
  regressions vs MCP        69         0
  natural micro/macro  94.0/88.9  94.0/88.9   0 queries better, 0 worse
  artificial            89.9/94.5  91.2/95.4  376 better, 1 worse
  artificial #1 exact     73.8%     77.6%

Zero: the site ranker now matches or beats rankEntries on every one of the 3,657
agent-shaped queries.

One judgement call for review: "default" now returns async-logger's `_default`
first, which the user previously said was less useful than what it returned
before. No natural query moved, so nothing real regressed, but the behaviour is
deliberate and worth a second opinion.

Author: Mikhus

src/_shared/js/search.js

@@ -412,6 +412,11 @@
     return folded.replace(/[^a-z0-9]/g, "");
   }
 
+  // Leading sigils a caller never types: `$lte`, `_default`, `#private`.
+  function sigilless(segment) {
+    return segment.replace(/^[^a-z0-9]+/, "");
+  }
+
   function lastSegment(name) {
     var parts = String(name).split(".");
 
@@ -875,7 +880,12 @@
 
     if (lower === q.joined) {
       direct = W.exact;
-    } else if (fold(lastSegment(lower)) === q.joined) {
+    } else if (sigilless(fold(lastSegment(lower))) === q.joined) {
+      // Sigil-insensitive, because nobody types the sigil. `FilterInput.$lte` is exactly what
+      // lastSeg exists for — "send" for RedisQueue.send — but its last segment folds to `$lte`,
+      // which never equalled `lte`, so the record fell through to W.substring and ranked 27th
+      // behind 26 titles that merely contain those letters inside aLTErnative and fiLTEr.
+      // Same for `_default`, `$gte`, `$in`: the sigil is TypeScript's, not the reader's.
       direct = W.lastSeg;
     } else if (lower.indexOf(q.joined) === 0) {
       direct = W.prefix;