[codex] trace pre-sampling skill and persistence latency

[rphilizaire-openai]

Summary

• Add trace spans around skill listing, selection, catalog rendering, and prompt loading.
• Trace extension and legacy skill injection along with connector/session bookkeeping.
• Trace conversation recording through rollout persistence, thread-store append, and local writer flush.
• Record bounded item, input, connector, catalog, and skill counts on the relevant spans.

Why

run_turn traces had a substantial pre-sampling attribution gap after skill/plugin preparation and before run_sampling_request. Skill context recording and the synchronous rollout flush were covered only by the outer turn span, making it impossible to distinguish skill work, event emission, and persistence latency.

This change uses function-level #[tracing::instrument] attributes only and does not change runtime behavior.

Validation

• just test -p codex-core-skills -p codex-skills-extension -p codex-thread-store — 196 passed.
• just test -p codex-core suite::skills::user_turn_includes_skill_instructions — passed.
• Broader affected-crate run compiled all four crates and ran 2,948 tests: 2,920 passed; 28 unrelated sandbox/timing-sensitive integration tests failed, including Seatbelt permission errors and websocket/mock timeouts.
• Rust formatting completed and git diff --check passed. The aggregate formatter also attempted unavailable uv/dotslash formatters and could not rewrite the parent justfile under the workspace sandbox.