Add --effort to review/adversarial-review subcommands (parity with task)

Summary

codex-companion.mjs task accepts both --model and --effort, but the review and adversarial-review subcommands accept only --model (valueOptions: ["base", "scope", "model", "cwd"]). Reviews therefore always run at whatever model_reasoning_effort is set in ~/.codex/config.toml — a machine-global file shared with the desktop app — with no per-call control.

Why it matters

Teams that drive reviews programmatically (stop hooks, CI-adjacent gates) may want review effort scoped per invocation without changing the global config that also governs interactive desktop sessions. The plumbing already exists: runAppServerTurn accepts an effort option and forwards it in turn/start; the adversarial-review path calls runAppServerTurn today, just without populating effort.

Suggested change

• Add "effort" to the review subcommands' valueOptions, validate with the existing normalizeReasoningEffort, and pass it through executeReviewRun → runAppServerTurn (the native review/start path may need its own handling or can reject the flag explicitly).

Related footgun worth closing in the same pass

parseArgs treats an unknown --flag as a positional, and review positionals join into the reviewer's focus text. So a caller who guesses --effort high today gets no error — the tokens are silently prepended to the review instructions. Erroring on unknown flags (or shipping the flag) would close that.

Observed on plugin 1.0.6 with codex-cli 0.147.0.

Start in codex-companion.mjs by tracing review and adversarial-review argument parsing, executeReviewRun, and the existing runAppServerTurn effort handling. Compare their options with task, including normalizeReasoningEffort and valueOptions. Done means per-invocation effort reaches the review path or is explicitly rejected for native review/start, and guessed unknown flags no longer become review focus text.