[v2] Remote execution end-to-end + horizontal scaling

[devzeebo]

Part of the Orchestrator v2 rebuild — a tear-down/rebuild in a new self-contained orchestrator-v2/ folder (Vite+ / vp, vp test + vitest-gwt, no hooks, single RPC/WebSocket runner interface, engine adapters copied wholesale). Completes #29 (runner split). See also #30, #31.

Problem

#29's headline: the orchestrator must dispatch to N runners and get results back so workload moves onto the runners and the system scales horizontally — replacing v1's one-thing-at-a-time in-process loop.

Proposed solution

Wire the pieces from the prior stories into an end-to-end remote-execution path:

• Orchestrator dispatches signed tasks to connected runners; each runner executes locally (including the local engine) and reports signed results.
• Tasks fan out across runners; the orchestrator collects results and calls complete/fail/pause on the source.
• Demonstrates the Task Agent (agent-3-task) running on a real runner.

Alternatives considered

• (Covered by [feat] Separate Agent Runner from Orchestrator #29 and the protocol/runner stories.) mTLS / OAuth / JWT / encryption are explicitly not used — signing only.

Area

Orchestrator v2 (completes #29)

Acceptance criteria

• With 2 runners (loopback or remote) + orchestrator: dispatching 4 tasks that each do ~50 ms of work completes in ~100 ms wall (parallel), not ~200 ms (serial).
• Tampered dispatch, tampered result, and expired timestamp are each rejected; the clean path verifies.
• An LLM Task Agent task runs on a runner end-to-end.

Dependencies

• [v2] Thin orchestrator
• [v2] Runner package
• [v2] Task Agent: LLM execution as a script

Verification

vp test + vitest-gwt integration: parallel wall-time across runners; signature/replay rejection; Task Agent on a runner.