fix(parser): handle nested command tags in skill bodies (#1340)

[averymkeller83-hub]

Summary

• Fixes UX: Skill instruction content visible in chat when invoking skills #1340: when a slash-command's skill body contains literal <command-message> / <command-name> examples in its markdown, the SDK wrapper ends up with nested closing tags, and the parser falls through to kind: 'text' — rendering the entire skill body as plain text in the chat.
• Two changes in parseLocalCommandMessage.ts:
  1. COMMAND_MESSAGE_RE becomes greedy so the whole outer block (including any nested example tags) is consumed in one pass.
  2. <command-message> blocks are stripped before the name/args extractors run, so a nested <command-name>/compact</command-name> example inside the wrapper can't be mistaken for the real command name.
• Adds one regression test reproducing the /go skill scenario from the issue.

Why both changes are needed

Either fix alone is insufficient:

• Greedy alone → stripping is fixed, but text.match(COMMAND_NAME_RE) still finds the first <command-name> tag in the original input, which is the nested example (/compact) rather than the outer wrapper's real command (/go).
• Reorder alone → without greedy, the lazy match still stops at the inner closer, leaving the outer </command-message> behind.

The community comment on #1340 caught the lazy/greedy half but missed the extraction-order half; this PR addresses both.

Test plan

• pnpm vitest run sources/components/parseLocalCommandMessage.spec.ts — 15/15 tests pass (14 existing + 1 new regression test).
• pnpm typecheck — clean.
• All existing test cases (caveat, no-arg, with-args, empty-args, mixed-content, plain-text, echo detection) still pass — no behavior change for non-nested inputs.

🧙 Built with WOZCODE