#126: Design the native-agent harness comparison contract

[justsml]

Split from the Wayfinder mega PR #137.

Refs #126

Scope

• Define a fair Codex, Claude Code, and OpenCode comparison contract with attribution and honesty boundaries.

Review notes

This PR is intentionally ticket-scoped. It does not close the referenced issue unless the issue is already complete; CI provides the final integration check.

[justsml]

Council gate passed after repair. Four-lens Council review and the independent Spec review found the contract complete; Standards review required GPT-5.6 Sol’s initial native Codex arm to use the documented reasoning tier, now corrected while retaining OpenCode’s catalog-supported lowest tier. The PR changes only this dated research document. Type Check passed on the latest head and Build/Eval Gates passed on the immediately preceding doc-only head. Test/Biome/audit failures are unchanged current-main baseline failures.

[justsml]

Correction to the prior Council note: the native Codex GPT-5.6 Sol arm now starts with model_reasoning_effort=none, as required when a documented thinking-off control exists. The OpenCode bridge retains its catalog-supported lowest available tier.