grill-perf: allow thinking/reasoning control in portable-chat-v1 (chat_template_kwargs)

[plotarmordev]

Problem

portable-chat-v1 cannot send chat_template_kwargs, so on GLM-5.3-Flash (thinking on by default) a 64-token capped structured prompt produced only reasoning tokens: no answer-text timestamp in any of 18 requests, and the A/A receipt in docs/performance/2026-09-09-recipe-smoke.md is effectively a prose measurement. A hand probe with max_tokens: 1 and thinking on produced no content delta at all and failed.

Evidence

• docs/performance/2026-09-09-recipe-smoke.json: first_answer_text_us null on all measured lanes.
• Live probe fix that made the run work: adding "chat_template_kwargs": {"enable_thinking": false} to the request body.

Proposal

Workload-level request.thinking: "on" | "off" | null (null = don't send), serialized as chat_template_kwargs.enable_thinking for vLLM-style servers, recorded in plan.json. Keep portable-chat-v1 strict otherwise. Optional generic request.extra_body (closed map of scalars) recorded in the plan would cover reasoning_effort too.

[plotarmordev]

Resolved in PR #19. The new explicit thinking_control kind vllm-enable-thinking-v1 sends enable_thinking, while legacy thinking keeps its exact existing wire contract. Conflicting declarations are rejected. portable-chat-v1 stays strict; use the explicit vLLM profile for this provider extension. No generic extra_body or automatic key inference. CLI fixtures verify emission, identity and conflicts; actual template compliance remains an observed provider property, not guaranteed by sending the field.