Engineering lifecycle follow-ups from the #447 review

[TheAmericanMaker]

Non-blocking findings from the Fable re-review of #447 (attempt lifecycle on the MCP surface). None of them weakens a gate; each one is either a disclosure gap or a loop that can't make progress.

1. request_acceptance asks again after a replayed receipt.
   • An agent with write access to approvals/ copies a genuine approval under a new id. Both copies then read receipt-replayed, and traverse stops for an operator (correct).
   • request_acceptance instead discloses the pair and asks the person again. An accept mints a second genuine approval for the same candidate.
   • Fix: when a non-genuine approval's receipt names a request that exists in storage (replayed or expired), stop the way traverse does rather than asking again.
   • The docs no longer claim uniqueness (attempt-lifecycle.md, feat: attempt lifecycle on the MCP surface (start_attempt, capture_candidate, record_review) #447).
2. Gate limitations don't reach the person. buildAcceptanceRequest builds the presentation from standardLimitations, the re-read, and ignored approvals. It never includes gate.limitations. For example, 'the change has no stored brief.md' reaches gate callers but never the human deciding. Fix: merge gate.limitations into the presentation.
3. A planned change with no slices is stuck.
   • A crash between plan's change CAS and its slice writes leaves a change planned with zero slices.
   • Traverse then advises plan-slices, which plan refuses ('slices are planned once, from draft'). No action recovers it.
   • Fix: allow plan on a planned change that has zero slices, or have traverse emit blocked-needs-operator.

Refs #409, #447.