feat(0.15.0): cross-worker sandbox-reconnect durability

[drewstone]

Summary

A 15-minute agentic sandbox turn must survive the Cloudflare worker isolate dying mid-turn (deploy roll, CPU limit, OOM). runDurableTurn already replays a completed turn, but an interrupted turn re-runs from the top — the producer's streamPrompt generator died with the isolate.

The Tangle sandbox container is orchestrator-managed and outlives the worker. This PR adds runReconnectableTurn: it checkpoints a RunHandle at turn start so a fresh worker re-attaches to the in-flight sandbox run instead of re-prompting.

• RunHandle — { kind: 'sandbox' | 'tcloud', sandboxId?, sessionId?, runId?, status, cursor? }. A pointer to a substrate run that outlives the isolate.
• runReconnectableTurn — three resolution paths on a retry: replayed (turn already finished — cached text replays), reconnected (a running handle survived — calls the product's reconnect(handle) callback), rerun/fresh (no reconnectable handle — produces live).
• reconnect(handle) is product-supplied substrate glue. Sandbox products wire the SDK's event-replay endpoint (GET {runtimeUrl}/agents/run/{runId}/events?lastEventId={cursor}); tcloud products omit it and fall through to a clean re-run.
• Storage: the handle is checkpointed as a completed step at index 0; the turn runs at index 1. Reuses the existing completeStep JSON-result path with zero schema change — a completed step is the only shape startOrResume returns to a retry, and the handle must be readable while the turn step is still running. A new durable_steps column would force a migration across all three stores plus a new store method.

This is a thin handle registry, not a second durable-execution framework — the sandbox runtime is the durable engine; agent-runtime just remembers the pointer.

Spike findings (@tangle-network/sandbox@0.1.2)

Cross-worker attach is feasible. streamPrompt's reconnect uses executionId (run id, carried on the execution.started SSE frame's data) + lastEventId (the SSE id: cursor). The runtime exposes GET {runtimeUrl}/agents/run/{executionId}/events?lastEventId={cursor}&format=sse, reachable from any process via the public SandboxConnection.runtimeUrl + authToken. The SDK does not expose a one-call resumeRun(executionId) — its reconnect loop is closure-local — so the raw replay fetch is product-owned, which is exactly why reconnect is a product-supplied callback.

Test plan

• pnpm typecheck passes
• pnpm test — 231/231 pass (18 new in run-handle.test.ts)
• New tests run across the InMemory / FileSystem / D1-over-sqlite store matrix
• Covered: fresh turn registers a handle; retry with a running handle calls reconnect not produce; completed handle replays; running handle with no reconnect falls through to re-run; reconnect-stream failure fails the run (error not swallowed); register advances the persisted cursor
• Confirmed the FileSystem-store concurrent-write race fix (handle writes drained before the turn-step write) is stable across 5 consecutive runs

[tangletools]

Verified after bringing the branch current with main (merged 9 commits incl. the model-resolution primitive, no conflicts). runReconnectableTurn + RunHandle: checkpoints a run pointer at turn index 0, a fresh worker re-attaches to the in-flight sandbox run (replayed/reconnected/rerun/fresh) instead of re-prompting — reuses the existing completeStep path, zero schema change. Full gate on the merged tree: typecheck 0, 254 tests green (18 sandbox-reconnect across the InMemory/FileSystem/D1 store matrix), build success, biome clean. CI green.