feat(agent-interface): bind exact candidate experiments

[drewstone]

Problem

Evaluation, approval, and execution could describe the same candidate with different task, grader, model, workspace, and receipt identities. That made an apparent improvement impossible to prove as the exact approved run.

Solution

• define one immutable benchmark task and experiment shared by evaluation and Runtime
• bind execution plans to instruction delivery, resource limits, model grants, workspace identity, candidate bundles, and graders
• require unique scenario, execution, run, receipt, and evidence identities
• derive cost and latency from exact receipts and reject contradictory results
• keep the greenfield candidate contract unversioned

Verification

• pnpm check-types
• pnpm test: 551/551 passed across 8 packages
• pnpm build
• pnpm --filter @tangle-network/agent-interface test: 120/120 passed after rebasing onto main
• git merge-tree --write-tree origin/main HEAD
• git diff origin/main...HEAD --check

[tangletools]

✅ Auto-approved drewstone PR — e7bf8632

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-16T05:39:21Z}