feat(studio): consume eval-runner failure-capture fields

[chubes4]

Summary

Pairs with Automattic/studio#3330, which adds two new fields to the eval-runner result JSON:

• result.error: string | null — set when an exception is caught inside the message loop (auth blip, MCP crash, network error, SDK throw, etc.)
• result.timedOut: boolean — set when the timeout callback fires before query.interrupt()

This PR teaches the Studio bench scripts to consume both fields.

Why

Today the bench gate is result.success === true && !result.error. On current trunk Studio, result.error is undefined outside the rare top-level catch path, so the AND degenerates to result.success === true. Timeouts are indistinguishable from clean model failures — both surface as success: false with no other differentiator.

After Studio#3330, the eval-runner emits both fields consistently, and the bench can:

• Distinguish timeouts from agent failures. A timed-out 5-minute run and a model that gave up at turn 3 today look identical in metrics. With this PR, timed_out is a separate metric.
• Distinguish runner-side exceptions from agent failures. A run where the model finished cleanly but the runner caught a recoverable exception today degrades to a stripped-down { success: false, error } JSON, losing all timings and tool history. After Studio#3330 the structured result is preserved with error set, and this PR's agent_runner_error metric makes it visible.
• Show the actual failure mode in error messages. studio-agent-runtime and studio-agent-site-info now print timed out after Nms / exception: ... / exit=N instead of always falling back to exit=N.

Diff

Three bench files, ~30 lines:

• studio-agent-runtime.bench.mjs and studio-agent-site-info.bench.mjs: failure-detail message branches on result.timedOut and result.error. Lifts the timeout literal into a named constant so the message stays in sync with the actual budget.
• studio-agent-site-build.bench.mjs: agentSucceeded gate now requires !result.timedOut. Adds timed_out and agent_runner_error metrics so timeout regressions are visible separately from agent failures.

Backwards compatibility

On Studio versions older than #3330, both result.error and result.timedOut are nullish (or absent). The new gates degrade safely: agentTimedOut resolves to false, !result.error resolves to true, and the bench behaves exactly as it did before. New metrics emit 0 instead of misleading values.

This means the rig PR can land independently of Studio#3330 — it's strictly additive on old Studio, fully active on new Studio.

Refs

• Automattic/studio#3330 — the upstream eval-runner change this consumes
• Automattic/studio#3262 — the umbrella ask for structured eval diagnostics
• Automattic/studio#3273 — the previous installment (firstToolError, toolEvents, phaseTimingsMs)

AI assistance

• AI assistance: Yes
• Tool(s): Claude Code (Sonnet 4.5)
• Used for: Drafted the gate-tightening, named-constant refactor, and failure-detail messages under Chris's direction.