v1.3.4

Patch. Lane-quality refinements and a packaging cleanup. No change to the frozen 1.0 contract.

• The research lane runs on opus and pins the version it answers for. Research is a silent-failure
  lane: a wrong research answer becomes the contract the implementation and its test are built against,
  so no later review catches it — every downstream artifact inherits the same premise. It is now on the
  strongest tier (like the review lanes), and its Method adds an explicit step — confirm each external
  source applies to the version actually in use in this repo (from the lockfile/manifest), and say so when
  a doc pre- or post-dates it, since sources disagreeing because one is stale is the common case.
• The review lane attacks the tests by default. A green suite is evidence only if its tests can fail.
  The code-reviewer lane now mutates the source behind new/safety-critical tests and confirms a test
  catches each mutation — reporting which safety properties are actually pinned versus merely green —
  instead of that check depending on the operator remembering to ask for it.
• Baton is a pure skill. Removed the plugin manifest, which never actually registered the skill (it
  declared no discoverable skills path) and would have forced a /baton:baton command. Baton is
  distributed as a skill — that is how the marketplace syncs it — so one identity, one /baton. Install
  docs corrected, with an explicit note that baton is Claude-Code-specific: its lanes need the Agent tool,
  background tasks, worktree isolation, and plan mode, which do not exist in claude.ai, Claude Desktop, or
  generic MCP clients (it loads there but the lanes do nothing).