feat(task): task-local thinking effort state, per-request override, and adaptive effort envelope (DTE series 2/5)

[easonLiangWorldedtech]

What

Introduce a task-local runtimeThinkingEffort override for reasoning effort, resolved per API request (next-request semantics), passed through metadata.reasoningEffort, and emitted as the adaptive-model effort envelope (output_config.effort) where applicable. This is the state + plumbing PR.

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 2/5 — standalone, based on main.

Design

Resolution order (strongest first):

1. task.runtimeThinkingEffort (task-local; written by the model tool in PR 3 and the user control in PR 4)
2. settings reasoningEffort
3. model default

• Task-local state is discarded at task end and never written to Settings.
• setRuntimeThinkingEffort(effort, source) is the single write path (source: model tool / user).
• ApiHandlerCreateMessageMetadata.reasoningEffort? — per-request override channel, same pattern as the abort signal (metadata read at request build time) and the existing Bedrock metadata.thinking precedent.
• Adaptive Anthropic class (supportsReasoningBinary models): when the effective effort resolves, the request carries output_config: { effort } clamped to [low, medium, high, xhigh, max] (soft-guidance envelope). The Bedrock handler already sends a hardcoded xhigh envelope — left unchanged (documented follow-up).

Acceptance criteria

• An override takes effect on the next API request (no mid-stream effect) for all providers with per-request effort support
• Adaptive-class Anthropic requests carry output_config.effort when effort resolves; omitted otherwise (no 400s)
• State resets at task end; persisted Settings value is never modified
• Tests: state set/get/reset, metadata precedence over settings, envelope presence/absence; 100% patch coverage; <= 1000 diff lines