Port loop/ + crew/ + spec-queue/ to a plugin (P1.5), with claim fencing #128

Parent: Phase 1a item P1.5 (3.9k LOC; rewrite expected — loop.ts is 105 KB and its
hook points are gone in opencode 2). Depends on plugin-repo-layout and slots after
port-peer-plugin (delegation ledger precedes loop completion messaging).

Why

The autonomous loop is the platform's engine: it claims ready work, runs it in isolated
worktrees, and hands off on completion (charter §8.5–8.8). In the fork this lives in a
105 KB loop.ts hooking upstream internals that opencode 2 has removed — so this port
is a rebuild against the plugin surfaces, guided by the existing spec suite
(crew-loop, swarm-coordinator, loop-done-handoff, escalate-before-idle,
loop-spec-queue, fix-queue-branch-awareness) and the landed behavior:
loop-completion-contract, the --auto flag and its Phase 5–6 tests
(52516a57eb, d5cf928208, 7b31baa46f, c509c05cb4, 70d6c7c41a, 1ecc4db407,
a97d9b7ec4), and crew claims as shared files (813393f21a).

What Changes

1. Rebuild the loop engine as the skein-loop-crew plugin package: spec-queue-driven
   work claims, execution, and completion handoff, keeping the landed behaviors above
   observable (drained vs found-nothing distinction, aggregated report, fenced auto
   runs, claim/release for cancellation).
2. Add claim fencing (charter §8.5): lease token, expiry, heartbeat, and a monotonic
   ownership generation so an expired worker cannot continue publishing stale work.
   Before reassignment, reconcile actual process/branch state — a stale heartbeat does
   not prove execution is stopped.
3. Register the skein worktree implementation via worktree.transform; worktree naming
   stays linked to work item + execution ID; no auto-delete of unmerged work (§8.6).
4. Completion handoff (§8.7) drives the next assignment from the workflow engine's own
   turn (record outcome → validate evidence → unlock downstream → assign ready work),
   never from an unbounded prompt loop (§23).
5. Role/capacity routing decisions stay llama-skein-tenant and policy-driven
   (per-model-placement-controls carries). The scheduling loop performs no LLM calls
   (§23).

Non-goals

• No global control-plane duties (charter §9.1 — skein-control-plane owns cross-repo).
• No new work-source authority: spec-queue remains the only claimable-work source until
  the P1.9/P1.5 ready-work contract lands.

Risks

• Rewrite scope is the known schedule risk (K2); ship the engine behind the landed
  behavior tests, growing coverage rather than porting the 105 KB blob verbatim.
• Claims data must survive plugin reloads — the shared-file claims store carries over;
  fencing tokens must be introduced atomically with adoption of the shared store.

Impact

skein-loop-crew plugin: spec queue + claims + fencing + handoff + worktree transform.
This is where crew autonomy becomes durable; unblocks autonomous-swarm-epic work and
the soak phase.