fix(kiro): map GPT-5.6 reasoning effort fields

[SemonCat]

Summary

• follow up fix(kiro): map GPT reasoning effort fields #2701's GPT-5.6 reasoning.effort mapping through the complete Responses-to-Kiro flow
• preserve Responses API reasoning.effort through the OpenAI intermediate
• retain Claude thinking plus output_config.effort behavior and legacy-model guards
• avoid GPT legacy thinking prompt tags when a supported native effort is emitted
• keep legacy fallback for unsupported effort values and explicit disable semantics

Why

Kiro's effort path is model-schema-specific: GPT models use reasoning.effort, while Claude models use output_config.effort.

This is a follow-up to #2701, which established the core GPT-5.6 structured effort mapping and was confirmed merged by the maintainer. This PR carries that flow through the actual Hermes codex_responses path and tightens the remaining edge cases.

The remaining gap after #2701 is end-to-end transport: the standard Responses API reasoning.effort is removed during Responses-to-OpenAI conversion unless it is copied into the OpenAI intermediate first. The base flow can also emit legacy thinking tags alongside the native GPT field and normalize xhigh/max too aggressively. This follow-up addresses those cases while preserving the older fallback behavior for unsupported values.

Official Kiro effort documentation: https://kiro.dev/docs/cli/chat/effort/

Resulting behavior

• GPT-5.6 low, medium, high, and xhigh use the native reasoning.effort field.
• max maps to GPT wire value xhigh.
• Supported native GPT efforts do not also receive legacy thinking tags.
• Unsupported auto, minimal, and ultra retain the legacy thinking fallback.
• none, off, and disabled remain intentionally disabled.
• Claude adaptive effort continues to use thinking plus output_config.effort.

Verification

• focused Vitest: 65 passed, 2 expected-fail baseline cases
• ESLint on all seven touched source and test files
• production npm run build
• git diff --check
• direct live image evidence: Responses-to-Kiro traffic reports THINK:high / THINK:medium with the fix(kiro): map GPT-5.6 reasoning effort fields #2707 patch stack
• structured code review: no actionable findings

Repository-wide lint and the broad translator snapshot suite contain unrelated upstream baseline failures; the focused suite covers the changed request paths and fallback semantics.