[Bug] /review high effort silently increased to 10× token cost via multi-agent fan-out

**Bug Description**
Title: /review at high effort silently became ~10× more expensive (single reviewer → 14-agent fan-out) in mid-August '26

Body:

The `/review` skill was aliased to `/code-review`and reimplemented into a multi-agent fan-out at higher effort levels, but the change is invisible at the call site: the same /review at the same high effort now costs roughly an order of magnitude more tokens and wall-clock than days earlier. Effort labels should mean roughly the same work across builds; this one changed under the label with no signal.

- Env: CLI 2.1.237 (Homebrew), model Claude Fable 5, effort high (long-standing default).
- Expected: /review at high costs about what it always did — a single read-only reviewer pass.
- Actual: for a normal +834/-55 PR, one /review high spawned ~14 subagents (~8 code-review finder agent () + ~6 code-review VERIFIER), consumed ~66M tokens (≈59M cache-read), took ~23 min, and was ~20% of a weekly-capped plan for a single review.
- Evidence it's a build change, not usage: the "code-review finder agent ()" / "code-review VERIFIER" prompt signatures appear for the first time on 08-21; before that, reviews used one "stringent code reviewer … READ-ONLY" subagent. Same effort label, ~1 agent → ~14.
- Suggestions: keep effort-label cost stable across builds or surface it (e.g. "this review will fan out to N agents"); consider gating the fleet to xhigh/max or opt-in; document the apparent model-gating (fleet on for Sonnet-5 high+, off for Opus/default) — the model-table vs observed-Fable-fan-out mismatch is confusing.

**Environment Info**
- Platform: darwin
- Terminal: Apple_Terminal
- Version: 2.1.237
- Feedback ID: 3cca2c35-06c0-49ea-931d-1aff32b6ca05

**Errors**
```json
[]
```

Start at the /review alias and /code-review implementation, then inspect the effort-level fan-out and model-table paths mentioned in the report. Reproduce with CLI 2.1.237 at high effort and compare agent count, token usage, and wall-clock against the single-reviewer behavior. Done means cost or agent expansion is stable or clearly surfaced, with model gating documented or corrected.

Scritto dal modello di indicizzazione a partire dal testo della issue.