feat(ship): run independent decompose-and-implement sub-tasks concurrently

[BorisTyshkevich]

Motivation

decompose-and-implement-loop.workflow.mjs always ran a High-risk+Large unit's
sub-tasks strictly sequentially — one coding agent at a time on the same branch —
even when a sub-task's own dependsOn declared it had nothing to do with another.
On a large plan (#630 phase 7, 7 sub-tasks) that costs real wall-clock time (~107
minutes) even when several sub-tasks are genuinely independent.

What changed

• Sub-tasks are grouped into dependency-ordered waves from their existing
  dependsOn graph (a plain topological pass, no new field). A wave of one
  sub-task runs exactly as before, directly on the unit branch. A wave of more
  than one runs those sub-tasks concurrently, each in its own
  isolation: "worktree" on its own wip/<unit>-<id> branch, since none of them
  can see another's edits until merged.
• After a concurrent wave, one integration agent merges every branch into the
  unit branch in that wave's order (git merge --no-ff) and gates the merged
  result before the next wave starts. A merge conflict there means the sub-tasks'
  declared file scopes were not actually disjoint — treated as a real planning
  defect (error status), never forced through.
• The decomposition prompt now tells Fable to favor adding a dependsOn edge
  whenever it isn't fully confident two sub-tasks are safe to run at the same
  time — dependsOn is the only signal driving concurrency, so a plan with
  sparse dependency information still comes out mostly or fully sequential
  (today's behavior unchanged), and only a genuinely independent plan gets real
  concurrency. This matches the skill's existing "when in doubt, serialize"
  principle for wave-planning at the unit level, applied one level down to
  sub-tasks.
• SKILL.md step 2.3 updated to describe the new wave/integrate mechanics.

Verification

No dedicated test harness exists for .workflow.mjs scripts elsewhere in this
skill (none of the other reference scripts have one either) — verified by careful
review of the control flow, in particular:

• phase('Implement') is re-asserted at the top of every wave iteration so a
  prior wave's phase('Integrate') label doesn't leak into the next wave's
  solo-sub-task agent call.
• The PARALLEL_TASK_SCHEMA/INTEGRATE_SCHEMA additions are additive; the
  single-sub-task-per-wave path (today's common case) is otherwise byte-for-byte
  the original prompt/logic, so existing behavior for a fully-sequential plan is
  unchanged.

Full repository gate green (check:types/check:arch/check:schemas/
check:examples/npm test/npm run build) — this change touches only
skills/**, outside the coverage-gated src/** tree.

Co-Authored-By: Claude Sonnet 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01LwFPT465eDJqYcRa8HGNLz