fix(runtime): normalize Kimi temperature in worker-side helper calls

[drewstone]

Problem

The agent-lab powered binary follow-up found that Kimi author/reproducer calls can be normalized in the lab runner, but worker-side helper calls inside generated/runtime strategies still hit a separate chat path that sends temperature below 1.

Observed in agent-lab run runs/2026-06-18/powered-binary-followup/kimi-k27-binary-n12p3-invalid.json:

• model: kimi-k2.7-code
• cell: kimi-k27-binary-n12p3
• status: stopped invalid after 12/12 gen0 rows and 3/12 gen1 rows
• trace: runs/2026-06-18/powered-binary-followup/kimi-k27-binary-n12p3-trace/events.jsonl
• error count: 6 invalid temperature: only 1 is allowed for this model events

Example trace event shape:

LLM call 400: {"error":{"message":"invalid temperature: only 1 is allowed for this model" ...}}

Why it matters

This makes Kimi unusable for certifiable strategy-evolution experiments even when the runner config sets WORKER_TEMPERATURE=1 and wraps the author ChatClient. The failure appears in worker-side helper calls such as critique/analyst paths used by generated strategies, not just top-level worker calls.

Expected fix

Normalize or retry temperature-rejected Kimi requests wherever runtime worker/helper calls bind agent-eval chat clients, matching the existing router-client behavior that retries temperature errors at temperature=1.

Guardrail

Do not hide these as scored model failures. Either retry with the provider-required temperature or surface a structured infra/model-adapter error so evals can mark the cell invalid instead of treating it as a bad policy.