scan: refuse a candidate whose modifier opens on - or _

CHANGES.md

@@ -196,6 +196,10 @@
   which used to drop the rule and leave an element carrying the class with
   nothing at all. `bg-size-[foo]` no longer writes `background-size: auto`, a
   value the class never asked for (#761 through #771).
+- The source scanner draws the boundary Tailwind's extractor does: a candidate
+  whose `/modifier` opens on `-` or `_` is not one, so `tw file.html` no longer
+  writes rules for `group-hover/-2a:underline` that the author's own toolchain
+  never produces (#772).
 - A negated arbitrary length is `calc(<value> * -1)` on every family and in
   every unit, the spelling Tailwind writes. The sign was folded into the number
   from a unit table that each of margin, the inset sides and `text-indent` kept

lib/tools/source_scan.ml

@@ -147,6 +147,28 @@ let read_candidate d start =
   in
   loop start 0 0 None false
 
+(* Tailwind's extractor refuses a candidate whose [/modifier] opens on [-] or
+   [_], and refuses the whole candidate rather than truncating it: the CLI emits
+   nothing at all for [bg-red-500/-2]. Only a [/] the utility itself carries
+   counts, so one inside a bracket or a paren group - [aspect-[16/9]],
+   [bg-[url(a/_b)]] - is the value's, not a modifier's. *)
+let modifier_opens_badly candidate =
+  let len = String.length candidate in
+  let rec loop i bracket paren =
+    if i >= len then false
+    else
+      match candidate.[i] with
+      | '[' -> loop (i + 1) (bracket + 1) paren
+      | ']' when bracket > 0 -> loop (i + 1) (bracket - 1) paren
+      | '(' -> loop (i + 1) bracket (paren + 1)
+      | ')' when paren > 0 -> loop (i + 1) bracket (paren - 1)
+      | '/' when bracket = 0 && paren = 0 ->
+          (i + 1 < len && (candidate.[i + 1] = '-' || candidate.[i + 1] = '_'))
+          || loop (i + 1) bracket paren
+      | _ -> loop (i + 1) bracket paren
+  in
+  loop 0 0 0
+
 let candidates source =
   let d = decoded_utf_8 source in
   let len = Array.length d.chars in
@@ -161,7 +183,10 @@ let candidates source =
           String.sub source byte_start (byte_stop - byte_start)
           |> trim_candidate
         in
-        let acc = if candidate = "" then acc else candidate :: acc in
+        let acc =
+          if candidate = "" || modifier_opens_badly candidate then acc
+          else candidate :: acc
+        in
         loop stop acc
       else loop (i + 1) acc
     else loop (i + 1) acc

test/test_source_scan.ml

@@ -82,8 +82,44 @@ const f = fn(arg)|}
   Alcotest.(check bool)
     "a call is not a candidate" false (List.mem "fn(arg)" found)
 
+(* Tailwind's extractor refuses a candidate whose [/modifier] opens on [-] or
+   [_], and refuses the whole candidate rather than truncating it: measured
+   2026-09-13, the CLI emits nothing at all for [bg-red-500/-2], not the
+   truncated [bg-red-500]. tw's scanner read one where Tailwind reads none, so a
+   sheet tw built carried rules the author's own toolchain never writes. Only a
+   [/] the utility itself carries counts; one inside a bracket or a paren group
+   belongs to the value. *)
+let test_scan_modifier_boundary () =
+  let source =
+    {|<span class="group-hover/-2a:underline group-hover/_x:underline
+group-hover/edit:underline bg-red-500/-2 bg-red-500/50 aspect-[16/9]
+bg-[url(a/b.png)] bg-cyan-400/(--alpha) w-1/2">x</span>|}
+  in
+  let found = Tw_tools.Source_scan.candidates source in
+  let refused cls =
+    Alcotest.(check bool)
+      (cls ^ " is not a candidate")
+      false (List.mem cls found)
+  in
+  let scanned cls =
+    Alcotest.(check bool) (cls ^ " is a candidate") true (List.mem cls found)
+  in
+  refused "group-hover/-2a:underline";
+  refused "group-hover/_x:underline";
+  refused "bg-red-500/-2";
+  (* A modifier that opens on anything else is a candidate, and so is every [/]
+     the value carries. *)
+  scanned "group-hover/edit:underline";
+  scanned "bg-red-500/50";
+  scanned "aspect-[16/9]";
+  scanned "bg-[url(a/b.png)]";
+  scanned "bg-cyan-400/(--alpha)";
+  scanned "w-1/2"
+
 let tests =
   [
+    test_case "a modifier opening on - or _ is no candidate" `Quick
+      test_scan_modifier_boundary;
     test_case "unbalanced bracket stops at the newline" `Quick
       test_unbalanced_bracket_stops_at_newline;
     test_case "split whitespace" `Quick test_split_whitespace;