A task reopened by task_verify failed is redispatched onto its old, already-merged branch with no prompt pointer to the verification evidence; the worker re-submits the merged work

[ryanrborn]

Summary

A task reopened by task_verify failed was redispatched onto its old,
already-merged branch with no new commits, which produced an empty PR
(#1957 for bd-96mn8i) instead of new work fixing the reported defect.

• Dispatch.maybe_provision_worktree/2 now runs Worktree.reset_if_merged/4
  before create/3. When the task's last verification outcome was :failed,
  it forces a reset even though this repo's default GitHub merge method
  (squash) means the old branch tip is never a git-ancestor of main — plain
  merge-base ancestry alone never catches that case.
• The "stale branch" check is git merge-tree --write-tree <ref> <branch>
  producing the same tree as <ref> itself — i.e. merging the branch in
  would change nothing. This correctly fires right after a squash lands
  (any number of squashed commits), keeps firing even after unrelated PRs
  land on main afterward, and stops firing the moment new work is
  committed on the branch (so a later redispatch in the same re-work round,
  while verification_outcome is still :failed, doesn't throw away that
  new work).
• A resumed dispatch (opts[:resume]) skips the reset entirely, and a live
  worktree with uncommitted changes is never hard-reset even under
  force: true — both guard against arb resume losing in-progress work.
• The work prompt now includes a verification_evidence section right
  after Acceptance, under a heading stating the merged fix failed in
  production, whenever verification_outcome == :failed.

Went through 3 rounds of internal review; round 3's finding was that the
prior tree-equality check (git diff --quiet ref branch) only matched
immediately after the squash landed — once anything else merged into
main, the trees diverged and the stale branch was wrongly kept again.
Switched to git merge-tree --write-tree, which is robust to main
advancing.

Test plan

• mix test on worktree_test.exs, dispatch_test.exs,
  prompt_builder_test.exs — 249 tests, 0 failures.
• mix format --check-formatted on touched files — clean.
• mix credo on touched lib files — no issues.
• Full mix test (umbrella mix precommit) — failures are pre-existing,
  unrelated ProvisioningTest (hardcoded worktree path assertions) and
  arbiter_cli (live dev workspace not named "default") failures,
  reproduced independently of this change.
• New test reproduces the round-3 scenario directly: squash-merge a
  2-commit branch onto main, then push an unrelated commit to main
  before redispatching — force: true still resets.
• Existing test proves a round-2 commit on a force-reset branch survives
  a later redispatch while verification_outcome is still :failed.

References

bd-8ssxap

Closes #1959