fix(tools): exclude a preceding slash from the glued-doc-comment guard

DESIGN-NOTES.md

@@ -710,10 +710,11 @@ the damage was real and the stated consequence was not -- worth separating, beca
 what would have driven the wrong fix (a lint setting that does nothing).
 
 [tools/check-encoding.ps1](tools/check-encoding.ps1) now rejects `///` immediately following a non-space
-character at end of line, in `.rs` files. The pattern was checked for false positives across the whole
-repository before being adopted: there are none, because a legitimate doc comment is always preceded by
-whitespace or begins its line. The script's own description was widened from "encoding" to **text hygiene**,
-since neither this rule nor the control-character rule is an encoding fault; they live here because this is the
+character, in `.rs` files. The pattern was checked for false positives across the whole repository before being
+adopted: there are none, because a legitimate doc comment is always preceded by whitespace or begins its line.
+A preceding slash is excluded so a `////` banner comment is not flagged. The script's own description was
+widened from "encoding" to **text hygiene**, since neither this rule nor the control-character rule is an
+encoding fault; they live here because this is the
 check CI already runs over every tracked file.
 
 The first version of this guard was a **no-op**: it gated on a `$ext` variable that was not in scope inside the
@@ -722,6 +723,12 @@ that a guard has been written, observed to pass, and believed. The rule that cau
 guard you have only seen pass is untested.** Plant the defect it is meant to catch and watch it fail, then
 remove the defect and watch it pass. Both directions, every time.
 
+The second version was anchored at end of line (`\S///\s*$`), so it caught only the *bare* glued marker -- the
+rarer form. The typical splice glues a whole doc line onto the code (`let x = 1;/// The next thing`), and that
+passed. This is a narrower rule than "test the guard": the guard *was* tested, against the exact instance that
+motivated it, and the instance was not representative. **A guard written from one observed defect must be
+checked against the general shape of the defect, not only against the instance in hand.**
+
 ## Restoring a file with an old timestamp silently disables the rebuild
 
 Verifying a fix by planting the defect back requires *building* both states. Restoring the fixed file with

tools/check-encoding.ps1

@@ -154,17 +154,21 @@ foreach ($file in $files) {
     #    leaving its own `# Errors` section truncated or empty where the moved
     #    text had come from.
     #
-    #    `\S///` is the whole condition: a doc marker touching a non-space
-    #    character. The surrounding `^.*` and `.*$` the pattern used to carry are
-    #    not just redundant once the anchors are gone, they are the expensive
-    #    part -- `^.*` matches to end of line and then backtracks looking for the
-    #    marker, on every line of every file. Measured over 1.47 MB of this
-    #    repository's own Rust, dropping them cut a no-match scan (the case CI
-    #    runs on every clean build) from 872 ms to 288 ms across 20 passes. The
-    #    match index still identifies the line, because the marker and the
-    #    character it is glued to are on it.
+    #    A doc marker touching a non-space character is the whole condition. The
+    #    `^.*` and `.*$` the pattern used to carry are not merely redundant once
+    #    the anchors are gone, they are the expensive part -- `^.*` matches to
+    #    end of line and then backtracks looking for the marker, on every line of
+    #    every file. Measured over 1.47 MB of this repository's own Rust,
+    #    dropping them cut a no-match scan (the case CI runs on every clean
+    #    build) from 872 ms to 288 ms across 20 passes. The match index still
+    #    identifies the line, because the marker and the character it is glued to
+    #    are on it.
+    #
+    #    A preceding slash is excluded -- `[^\s/]` rather than `\S` -- so a
+    #    `////` banner comment is not flagged. `\S///` matched every one of them,
+    #    because the third slash of the banner is itself a non-space character.
     if ([System.IO.Path]::GetExtension($file) -eq '.rs') {
-        $glued = [regex]::Match($text, '\S///')
+        $glued = [regex]::Match($text, '[^\s/]///')
         if ($glued.Success) {
             $prefix = $text.Substring(0, $glued.Index)
             $line = ($prefix -split "`n").Count