Build the runtime BoardWorker seam

The board needs a runtime drain that follows the same seam discipline as the existing Orchestrator (Runtime orchestration & per-tenant control plane — the learning loop becomes a runnable system): an interface the runtime depends on, a DefaultBoardWorker that ships now (deterministic, offline, CI-green), and a reserved AgentSdkBoardWorker that is the only place @anthropic-ai/claude-agent-sdk would import — stubbed to keep CI offline, exactly like the reserved AgentSdkOrchestrator. This activates the work queue that Agentic "sprint board" architecture — a git-resident OKF task board worked by bounded, hook-governed Agent SDK loops says the reserved seam drains.

Shape (analogy to the Orchestrator seam)

• BoardWorker { name; claim(tenant, opts) -> ClaimedTask | null; execute(task) -> TaskResult }.
• DefaultBoardWorker — reads tenant.okfRepoPath task atoms, applies claim/lease (frontmatter timestamps), transitions deterministically, no network.
• AgentSdkBoardWorker — reserved stub; bounded (maxTurns + maxBudgetUsd), honors the board-level pause flag, writes the transition back, commits.
• drainBoard(tenant, opts) — claim → execute → commit, bounded (maxTasksPerRun, idempotency keys) and resumable, as a new board.ts plugged into index.ts with zero changes to loop.ts/orchestrator.ts/provision.ts.

Depends on

Design the OKF `task` concept type + seed the board (the task schema it reads). Pairs with the governance hook Build the PreToolUse claim/lease governance hook (claim enforcement) and assumes the compounding-merge guard from Make the learning loop compound (reconcile merge by id + confidence) (repeated agent writes must not clobber curation). The durable scheduler host (open question 1 of Agentic "sprint board" architecture — a git-resident OKF task board worked by bounded, hook-governed Agent SDK loops) is a separate deployment concern — this task delivers the bounded worker abstraction, not the harness.