efficient-fable: add guidance for staying unblocked during execution

[fabb]

Adds a "Staying Unblocked" section to the efficient-fable skill with three execution-time practices, and strengthens the existing stop-conditions guidance:

• Defer user questions until no other independent work remains. Every mid-plan stop stalls slices that could keep moving; batch open questions for the end. Exception: ask immediately when the answer gates every remaining slice or the next step is irreversible.
• Don't sit blocked on a running subagent either — run subagents in the background and spend the wait on orchestrator-side work (vetting earlier results, planning the next slice).
• Keep the build green even with incomplete changes — leave unfinished work stubbed or unwired rather than half-connected — so one blocked or stalled slice does not block the parallel ones. If a slice breaks the build, revert only its diff in the working copy.
• Stop-conditions bullet in Handoff Packets: make explicit that subagents never pause to wait on user input — they stop the task and report, keeping the orchestrator unblocked. (In Claude Code, Task-tool subagents cannot prompt the user anyway; this keeps the behavior explicit for other harnesses.)

Potential further improvement (not in this PR)

A natural companion: before the orchestrator finally does block on a user question, write the full current state to a HANDOFF.md in the repo root — per-task status, decisions made and why, dead ends, the open question — so work can resume from a fresh context. Overwrite it on each update, never commit it, delete it when the task completes. This is more about orchestrator resumability than delegation efficiency, so it may fit better as a separate skill; happy to open a follow-up PR either way if there's interest.