fix(bench): read current Pier candidate contract

[drewstone]

Problem

@tangle-network/agent-bench@0.4.0 reads the pre-#547 candidate layout, so every Runtime 0.105 Pier task fails before execution (plan.bundleDigest is absent).

Change

• extend the existing Python Pier bridge to read the signed run cell, embedded benchmark task/suite, and profile activation
• preserve fail-closed bindings for digests, coordinates, limits, model, container, retry policy, harness, and activated profile bytes
• accept only the fixed executable PATH already signed by Runtime
• publish agent-bench 0.4.1 and update its real failure/success proof for the current receipt layout

Checked before editing: the bridge already existed in bench/pier_agents; this updates it rather than adding another executor.

Proof

• real Pier pair: failure 0/1, exact repair 1/1, fresh-process recovery 1/1
• both arms: 0 model calls, 0 input/output tokens, $0
• 30/30 focused Python tests
• 61/61 agent-bench TypeScript files plus Pier bridge
• packed consumer: agent-bench 0.4.1 with Runtime 0.105.0 on TypeScript 5.9.3 and 6.0.3
• pnpm verify:bench
• clean merge tree against current main after fix(bench): isolate evaluator Git from candidate config #620

A genuine-timeout trace-reporting defect remains separate; it does not affect completed trials or this release.

[tangletools]

✅ Auto-approved drewstone PR — 1e83914c

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 · 2026-07-25T01:32:52Z}