feat(tools): set_thinking_effort native tool — model-driven per-turn thinking effort, no approval (DTE series 3/5)

[easonLiangWorldedtech]

What

Add the native tool set_thinking_effort(effort, reason) so the model can adjust its own thinking effort mid-task, with an in-chat one-line display. No approval gate (non-destructive, clamped to the model's capability, instantly undoable); guardrails replace approval: always notify, escalation cap (max 3 upward changes per task), oscillation detection, hard clamp.

Part of the Dynamic Thinking Effort series (5 PRs). Master design issue: easonLiangWorldedtech#28.

PR protocol (all series PRs): draft PR until CI is fully green -> switch to ready -> CodeRabbit bot review addressed in full -> 100% test coverage on patch lines. Each PR <= 1000 diff lines, based on latest upstream main (stacked on preceding series PRs where noted; rebased to pure main once dependencies merge).

Stacking: PR 3/5 — stacked on PR 1 (experiment flag) + PR 2 (task state).

Design

• Tool exposed only when the dynamicThinkingEffort experiment is on and the model supports per-request effort (gating precedents: generate_image / run_slash_command experiments). Tool list stays stable within a task (prompt-cache safety).
• Schema: effort (enum = the model's supportsReasoningEffort capability array) + reason (required, surfaced in the UI display).
• Executor: clamp vs capability -> guardrails (escalation cap, oscillation) -> task.setRuntimeThinkingEffort(effort, "model") -> one-line chat display following the switch_mode display pattern (including partial streaming).
• Dispatch wiring: presentAssistantMessage + NativeToolCallParser cases.

Acceptance criteria

• Model can switch effort mid-task; display shows new effort + reason; the next request carries the new value (verifiable in logs)
• Escalation cap + oscillation notify work; effort is always clamped to the capability array
• No approval gate; experiment off => tool absent; task tool list stable
• Tests: executor (cap, oscillation, clamp, no-approval), gating, dispatch/parsing; 100% patch coverage; <= 1000 diff lines