Adds the design spec for HarnessBench — a harness benchmark that runs the same prompt across multiple model/config variants and compares how each performs. The doc lives at tech-specs/2026-06-agentic/harnessbench.md, alongside the other agentic specs.

This PR is the design doc only — no implementation yet.

Scope (MVP)

A new harnessbench worker (orchestration + metrics, no UI) plus an internal comparison view in console.

Highlights

• Benchmark run = one fixed prompt × N config variants ("legs"). Each leg can override model / provider / thinking-level and any harness SendOptions (system_prompt, max_turns, output, functions).
• Read + orchestrate, not instrument — every metric comes from already-persisted data: per-message usage (tokens/cost), transcript timestamps, and OTel spans.
• Two-tier metrics — Tier-1 from the session transcript (wall time, turn count, tool calls, tokens, cost, final context size); Tier-2 trace enrichment (span count, avg span time), best-effort so the MVP has no hard trace-retention dependency.
• Durable, event-driven orchestration — fan out harness::send per leg, complete reactively via harness::turn-completed, with a sweep fallback for dropped events.
• Worker surface — harnessbench::create-run | status | get | list.

Out of scope (later layers on the same functions)

Correctness grading, multi-turn scenarios, the standalone harnessbench.iii.dev site, and CI/regression baselines.

Ref: https://linear.app/motia/issue/MOT-3534/harnessbench