Foreground cyberine run with a build fleet runs through the engine (isolated worktrees)

When a build fleet is in effect (routed, --fleet, --cheap, or CYBERINE_BUILD_FLEET) AND the cwd is inside a git repository, the foreground cyberine run no longer dispatches the raw compare pool against the live tree. It runs the SAME implement-task workflow --bg uses (design -> implement -> review) through WorkflowEngine + RunStore: every implementer works in its own detached git worktree under .claude/worktrees/, DiffJudge applies each candidate diff in a throwaway worktree and type-checks it, and the winner lands in run.json as steps[].winnerOutput for cyberine apply. The live tree is not touched.

cyberine run --fleet muse,claude "add a clamp(x, lo, hi) function to math.ts"
# stderr (live, tailed from the run's events.jsonl):
#   Dispatching task <uuid> through the engine (implement: fleet muse,claude; ...)
#   [run:start]: Workflow "implement-task" queued
#   [step:start] design
#   [step:done] design: winner: ...
#   [step:start] implement
#   [worker:start] implement/muse ...
#   [step:done] implement: ...
#   [step:implement] state=done winner=claude
# stdout:
#   runId:  <uuid>
#   state:  done
#   diff:   1 file(s), +6/-0 from claude (step implement)
#   apply:  cyberine apply <uuid8>
#   watch:  cyberine agents

cyberine run --fleet muse -p "..."     # prints the winning DIFF to stdout (exit 1 when none applied)
cyberine apply <uuid8>                  # lands it; --dry-run first if you like

• The run is persisted like a --bg run, so cyberine agents, cyberine apply, digest and land all work on it; --timeout sends the run a stop control message.
• A qa task, or a run with no build fleet, still uses the read-only compare pool exactly as before.
• Outside a git repository the engine path is skipped with a NOTE and workers edit the tree in place, as they did before this change -- there is nothing to diff against.
• The review step's skeptic fleet is cheap and read-only; a review worker erroring does not fail the run.