/codex:review rejects focus text — breaks interface parity with /codex:adversarial-review and blocks non-English model/effort entry

Describe the bug

/codex:review throws on any positional argument because validateNativeReviewRequest rejects non-empty focus text unconditionally (codex-companion.mjs:271-274): "does not support custom focus text. Retry with /codex:adversarial-review ...".

1. It breaks interface parity. /codex:adversarial-review accepts focus text after the flags; /codex:review rejects it wholesale. Two sibling review commands that should behave as mirrors differ in a way the user only discovers by hitting the error. The suggested escape hatch (adversarial-review) is not a substitute — it changes the review semantics (adversarial framing), so the user is forced to pick between "type a sentence" and "get the review I actually want."

2. It makes non-English model/effort entry unusable. The short aliases (spark, sol, terra, luna) and the latin-only --flag value grammar assume the user types in English/Latin script. A user who writes the model/effort value as a natural phrase in their own language has no accepted input path. They must know the exact flag syntax and the exact latin model id.

To reproduce

A Japanese-speaking user requesting a high-effort review on the sol model, typed the natural way (model → モデル, sol → ソル, effort → エフォート, xhigh → エックスハイ — Japanese phonetically transcribes everything into kana):

$ /codex:review モデル ソル エフォート エックスハイ
Error: `/codex:review` now maps directly to the built-in reviewer and does not support custom focus text.
         Retry with `/codex:adversarial-review ...`

The intent ("use model sol, effort xhigh") is unambiguous to a human, but the parser treats the non-Latin prose as forbidden focus text — and even if it didn't, there is no ソル alias to map back to sol.

Expected behavior

Not mutually exclusive:

1. Lenient positional parsing: if a positional argument can be interpreted as model/effort intent, accept it as such instead of rejecting the whole invocation.
2. Recognize model/effort intent regardless of position/script: scan raw arguments for model/effort intent in any language before the strict rejection, so /codex:review モデル ソル resolves to --model sol.
3. Parity: accept the same argument surface in both review commands, so the interface is uniform.

Environment

Plugin v1.0.6, codex-cli 0.144.4, macOS.

Related

• #333, #512 — focus-text re-tokenization (inverse problem: flags leaking into focus text; this issue is benign focus text rejected wholesale).

Start in codex-companion.mjs:271-274 and trace validateNativeReviewRequest, then compare its argument handling with /codex:adversarial-review. Reproduce the Japanese positional invocation from the issue and determine how model/effort intent is currently parsed. Done means the agreed positional or multilingual model/effort inputs no longer fail wholesale and the sibling commands expose the intended matching argument surface.