Add Codex backend to skill eval harness

[amd-mkarvir]

What changed

• Add --agent codex to the existing routing and behavior eval runner.
• Install only the skills under test in a temporary, isolated CODEX_HOME.
• Parse Codex JSONL for routing and behavior grading using observed CLI event shapes.
• Add an advisory Codex routing workflow that skips when OPENAI_API_KEY is unavailable.
• Reuse the routing plugin installation for preflight instead of installing it twice.
• Treat retryable Codex error events as non-terminal and stop on turn.failed.
• Add a sanitized live JSONL fixture and regression tests for parsing and process termination.

Compatibility

Claude remains the default backend. The existing required evals gate is unchanged. Codex results remain advisory while a baseline is established.

Validation

• 91 unit tests passed.
• 68 eval cases across 8 skills and 5 shared negatives passed structural validation.
• A live read-only plugin activation was captured with Codex CLI 0.150.0-alpha.12.2 and gpt-5.4.
• The live run represented activation as item.started / item.completed with a nested command_execution reading the installed SKILL.md; it did not emit skill_invocation.
• eval/fixtures/codex-exec-skill-read.jsonl preserves the live event and field structure with identifiers, paths, prose, output, and usage values sanitized.
• The temporary plugin and marketplace were removed after capture.
• git diff --check and the public-content credential scan passed.

Fork pull-request workflows do not receive repository secrets, so the advisory Codex job skips cleanly when OPENAI_API_KEY is unavailable.

[danielholanda]

We are significantly refactoring our test harness this week. My recommendation is to not do this work here now.