fix(agents): preserve extended reasoning effort levels

[Filip-Stastny-Philips]

Prerequisite

• PR fix(agents): recognize GPT-6 model variants #602 It establishes the base GPT-6 recognition, routing, and limit handling that this PR extends with reasoning-level preservation.
• or PR feat: add GPT-6 family support (luna, sol, astra, 6.1-sol) #582 (both overlap over the same issue, depends which one is merged first)

Why

GPT-5.6 and GPT-6 effort selections were not consistently preserved. Pi downgraded xhigh and max to high, while the CodeMie OpenCode configuration did not explicitly expose model-specific Responses variants.

What changed

• Forward all six CodeMie effort levels through the built-in codemie-code OpenCode wrapper.
• Model GPT-5.6 and GPT-6 Sol/Luna with native none, low, medium, high, xhigh, and max levels.
• Model GPT-6 Astra and GPT-6.1 Sol with native low, medium, high, xhigh, and max levels.
• Keep CodeMie's minimal as a compatibility alias to the lowest native level (none or low).
• Add GPT-6 Responses routing, context limits, and Pi reasoning detection.
• Preserve Pi xhigh and max mappings and add focused regressions.
• Keep the detailed plan/findings in local-only tmp/opencode-max-effort-level.md; it is excluded from the PR.

Validation

• Focused OpenCode, Pi, and CodeMie reasoning tests: 65 passed.
• Typecheck and changed-file lint pass.
• Commit hooks passed with the local secrets scan skipped because no container engine is installed.