fix(composer): send commands and skills the way the harness expands them

[Tryanks]

Summary

What the composer sends, before and after, per provider:

 detect_composer_trigger(text, cursor)
-  "/" at the start of any line opens the command menu
+  "/" opens the command menu only at byte 0 of the message
+  "/Users/me/notes.md" (a second "/" in the token) is a path: no menu
 menu_visible()
-  true whenever a trigger is active ("No matching command." + Enter does nothing)
+  true only when the menu has rows ("@" keeps its loading/empty state)
+  → an unmatched "/word" or "$WORD" is plain text and Enter submits it

 wire_text()  (runtime, before the provider sees the turn)
   image-only placeholder
+  native_skill_invocation: leading "$name" → "/name" when the provider lists
+    name as both a skill and a slash command (Claude); Codex keeps "$name"
-  ultrathink: "Ultrathink:\n" + text
+  ultrathink: "Ultrathink:\n" + text, or text + "\n\nultrathink" after a "/command"

 codex provider start
   skills/list → "$" menu
-  ~/.codex/prompts/*.md → "/" menu   (app-server never expanded these)

Why: Claude Code expands a command or skill only when /name is the first byte of the user text; anything else reaches the model as plain text and the model then calls the Skill tool itself. Codex's app-server expands $name in plain text itself and has no API for custom prompts (deprecated upstream).

Evidence

• Before: typing /Users/me/notes.md opened an empty "No matching command." menu and Enter did nothing; $pongskill to Claude and Ultrathink:\n/pongcmd both produced TOOL-USE: Skill (model-driven), and /pongprompt on Codex made the model cat the prompt file.
  After: same stream-json messages the runtime now produces, driven straight into the CLIs:

  claude  /pongcmd hello            → USER-TEXT: <command-message>pongcmd</command-message>… → PONGCMD
  claude  /pongskill                → USER-TEXT: <command-message>pongskill</command-message>… → PONGSKILL
  claude  /pongcmd hello\n\nultrathink → <command-args>hello\n\nultrathink</command-args> + "ultrathink" system-reminder in the transcript
  codex   Use $pongskill (text_elements: []) → agentMessage: PONGSKILL, no file read
  

  New tests, all passing with the rest of the workspace (889 run, 0 failed):

  tcode-ui      composer_trigger::tests::triggers_respect_token_boundaries_and_the_utf8_cursor
                composer::components::trigger_menu::tests::unmatched_command_and_skill_tokens_open_no_menu
  tcode-runtime app::tests::skill_mentions_use_the_providers_native_invocation
                app::tests::ultrathink_keyword_trails_a_slash_command
  

Merge Danger

Door: two-way

Pure client-side behaviour; reverting restores the old trigger rules and wire text.

Blast Radius: composer

Three visible changes for existing users: / on a second line no longer opens the menu; the "No matching command / No skills found" empty states are gone (an unmatched token is just text); Codex users lose ~/.codex/prompts entries from the / menu. A $skill sent to Claude now arrives as /skill, so the bubble shows $skill while the CLI runs the command.