parallel flattened queue for recursive --help resolution

Replace recursive depth-first help_resolve with BFS queue + fork-based
parallelism (up to num_cores workers). Workers marshal results back via
pipes; discovered subcommands are enqueued for the next wave.

Also fix usage positional extraction to match "USAGE" without colon
(Go/Cobra style), and skip empty-result check to consider positionals.

lib/parser.ml

@@ -415,18 +415,21 @@ let extract_usage_positionals text =
       else
         let t = String.trim lines_arr.(i) in
         let tlen = String.length t in
-        if tlen >= 6 then
-          let prefix = String.lowercase_ascii (String.sub t 0 6) in
-          if prefix = "usage:" then begin
-            let after = String.sub t 6 (tlen - 6) |> String.trim in
-            if String.length after > 0 then Some after
-            else if i + 1 < len then
-              (* Clap style: USAGE:\n  cmd [OPTIONS] PATTERN *)
-              let next = String.trim lines_arr.(i + 1) in
-              if String.length next > 0 then Some next else None
-            else None
-          end else go (i + 1)
-        else go (i + 1)
+        let lc = String.lowercase_ascii t in
+        if tlen >= 6 && String.sub lc 0 6 = "usage:" then begin
+          let after = String.sub t 6 (tlen - 6) |> String.trim in
+          if String.length after > 0 then Some after
+          else if i + 1 < len then
+            (* Clap style: USAGE:\n  cmd [OPTIONS] PATTERN *)
+            let next = String.trim lines_arr.(i + 1) in
+            if String.length next > 0 then Some next else None
+          else None
+        end else if lc = "usage" then begin
+          if i + 1 < len then
+            let next = String.trim lines_arr.(i + 1) in
+            if String.length next > 0 then Some next else None
+          else None
+        end else go (i + 1)
     in
     go 0
   in