fix(ci): switch tarpaulin to llvm engine to fix coverage flake

[emal-avala]

Summary

Intermittent Coverage failures hit ~half of PRs with:

```
memory allocation of 93825031283952 bytes failed
test query::tests::cancel_before_first_event_interrupts_cleanly ...
ERROR cargo_tarpaulin: ECHILD: No child processes
```

Root cause: tarpaulin's default ptrace engine races with tokio's child-process reaper on cancellation tests. A ptrace-interrupted `wait()` returns a bogus value, which tokio interprets as an enormous allocation request.

Fix: pass `--engine llvm` to use source-based coverage via LLVM instrumentation. No ptrace, no race. Tests run on an unmodified binary, so behavior matches the non-coverage Test job.

Verification

• Output format unchanged (cobertura XML) — Codecov upload works as-is
• `--engine llvm` is stable in cargo-tarpaulin since 0.28
• The flaky test itself is correct — it's the observer that was wrong

Evidence the flake exists

• `main` after feat(cli): add /rename to label the current session #154 merged: Coverage failed with this exact signature
• PR feat(cli): add /rename to label the current session #154 (merged): Coverage failed
• PR feat(cli): add /add-dir to track extra directories alongside cwd #155 (merged): Coverage passed — it's intermittent
• PRs feat(skills): add remember bundled skill #156 / feat(skills): add stuck bundled skill #157 / feat(skills): add simplify bundled skill #158 (currently open): all three fail with identical signature

Test plan

• YAML syntax check — file parses
• CI on this branch runs Coverage green with the new engine
• Open PRs (feat(skills): add remember bundled skill #156–feat(cli): add /usage for per-turn token timeline #160) pick up the fix after rebase and go green

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.