chore(deps): widen agent-eval to ^0.91 + bump agent-runtime ^0.50

[drewstone]

What

• @tangle-network/agent-eval: >=0.77.0 <0.80.0 → ^0.91.0
• @tangle-network/agent-runtime: ^0.44.0 → ^0.50.0

Why

The <0.80 ceiling on agent-eval blocked the fleet from a consistent agent-eval and risked a duplicate install of the substrate everyone depends on. 0.91 is what the rest of the fleet uses. After the bump, agent-eval resolves to a single 0.91.0 (agent-runtime 0.50 consumes it as a peer at the same version — no duplicate).

Substrate API change handled

agent-runtime 0.50 removed createFanoutVoteDriver from the /loops surface (it's gone from the package entirely, not renamed). multiHarnessResearcherFanout used it to build a "fanout of N harnesses, pick the best valid output" topology.

Reconstructed the identical semantics with the supported primitive: createDriver + a single-fanout-then-stop planner. The kernel round-robins agentRuns across the N branches and selects the winner across iterations via defaultSelectWinner — same behavior, current API. The driver's action type moved from the old 'pick-winner' | 'fail' to the kernel's DriverDecision ('continue' | 'done').

Two tests asserted the old driver's internal name ('fanout-vote') and decision strings ('pick-winner' / 'fail'); updated to the new createDriver contract ('dynamic' / 'done'). The winner-selection assertions — the actual behavior under test — are unchanged and still pass.

Verification

• pnpm install — agent-eval 0.91.0 (single install), agent-runtime 0.50.0
• pnpm run build — pass
• pnpm run typecheck — pass (0 errors)
• pnpm test — 110 passed / 5 skipped (live-network)
• pnpm run lint — pass (no new warnings)

[tangletools]

✅ Auto-approved PR — dcf27d24

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 · 2026-06-14T00:55:35Z}