refactor(bench): WorkerBackendType = SDK BackendType (A-converge step 1)

[drewstone]

A) converge cleanup — step 1

First, safe step of retiring the non-unified flows onto the one resolver / canonical SDK types.

WorkerBackendType was a hand-maintained literal copy (opencode|hermes|claude-code|codex|kimi-code|pi) — a strict subset of @tangle-network/sandbox's exported 14-value BackendType, drifting from the real harness set. Aliased to the SDK type: one source of truth, no behavior change (widening only). All bench backend selectors (experiment.ts, run.ts, run-benchmarks.ts, search-bench/run.mts) now reference the canonical type.

Typecheck clean.

Why this is step 1 of a staged converge (not the whole thing)

The remaining A) items are load-bearing and each is generic-loop + domain-glue, so they need individual tested PRs, not a big-bang into the kernel:

• runAgentic's shotExecutor/agenticRegistry (off-box) → resolve worker from createExecutorRegistry (keep both loop substrates; both resolve from the one registry).
• benchSolverRegistry (solve+judge leaf) → extract the shared off-box-completion path, keep the judge as a seam.
• agent-lab gym-agent off-box loop → superseded by feat(bench): generalized AgentProfile-coordinate optimizer on the sandbox surface #293's sandbox-surface worker once EOPS gym is box-routable.
• rsi.ts BACKEND-env → profile-driven harness (needs harness as a first-class profile field).

These converge onto #293's profile-coordinates + the createExecutorRegistry resolver.

[tangletools]

✅ Auto-approved PR — 33e5ebe2

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-14T14:02:32Z}