Shard managed Codex integration signal

[lilly-luo]

Hypothesis

Long managed Codex lanes increase exposure to hosted-runner shutdowns and make the affected scenario opaque. Splitting the lane across fresh, bounded runners should reduce that exposure and localize any hang.

Change

• move managed-config coverage into an independent signal-only workflow, so runner cancellation cannot cancel required e2e
• split Codex into disjoint workspace, discovery-policy, and remaining-fixture shards
• retain a single Claude signal shard
• keep unique artifacts and 25-minute per-shard bounds

Validation

• uv run pytest tests/test_integration_contract.py -q (4 passed)
• uv run ruff check tests/ scripts/run_integration.py
• uv run ruff format --check tests/ scripts/run_integration.py
• collection selectors: Codex workspace 1 test, discovery 12 tests, remaining fixtures 6 tests
• workflow YAML parsed successfully

This is one experimental CI restructuring option for comparison; the managed signal remains non-required.