dispatch prompt: EnterWorktree-first isolation; manual git worktree add stalls Claude bg agents on approval

[MJohnson459]

Rewrites the dispatch preamble's two branch blocks so the agent isolates with
its harness's own worktree mechanism first, and adds a repo-level permission
allow so a Claude dispatch cannot stall on the approval prompt either way.

Why

The blocks told every agent to isolate with git worktree add <path> -b <branch>. For a Claude Code background agent that is a trap: the harness
rejects file edits until its own EnterWorktree tool has run, and pointing
that tool at a pre-made worktree raises an interactive approval prompt a
headless session can never answer. On 2026-07-31 three of four dispatches
(#312, #313, #294) followed the instruction literally and froze about two
minutes in, sitting running for eleven hours on dead sessions; the one agent
that used name-based EnterWorktree instead (#314) finished.

What changed

• crates/voro/src/dispatch.rs: new BRANCH_ISOLATE_SENTENCE, shared by
  ASSIGNED_BRANCH_TEMPLATE and UNASSIGNED_BRANCH_TEMPLATE beside the
  existing register and rebase sentences so the three cannot drift. It names
  the harness mechanism first (EnterWorktree, "creates and enters a worktree
  under .claude/worktrees/ unprompted"), demotes git worktree add <path> -b <name> to the fallback for agents without one, and — since such a tool names
  the branch itself — spells out the git switch -c <name> that puts the work
  on the branch Voro tracks. The unassigned template now uses <name>
  throughout instead of mixing <name> and <branch>.
• Invariants held: never modify the primary checkout, never push, early
  registration via voro set --branch, and the stale-base rebase sentence are
  all unchanged and still shared.
• .claude/settings.json (new): allows EnterWorktree, so even a path-entry
  call cannot hang a headless job in this checkout.
• docs/DESIGN.md §11 worktree lifecycle: records that how the agent makes its
  worktree is its own business, why the harness mechanism is preferred, and
  that a nested worktree is immaterial downstream because open and cleanup
  both find it through git worktree list.

Verification

cargo test --workspace (542 tests) and cargo clippy --workspace --all-targets -- -D warnings both clean. New test
preamble_puts_the_harness_worktree_tool_ahead_of_git_worktree_add asserts,
for both the assigned and unassigned cases, that the prompt names
EnterWorktree, carries git switch -c <branch>, and mentions the harness
mechanism before the manual fallback. Both rendered preambles were printed in
full and read end to end. This task itself was done through the new flow —
EnterWorktree, then git switch -c task-320-enterworktree-dispatch — with no
approval prompt.

Branch task-320-enterworktree-dispatch is committed locally and not pushed.
Follow-up filed as #324 (nested worktrees show untracked in projects that do
not ignore .claude/worktrees/).