code-review skill: --level low still runs the full high-effort 8-angle pipeline

Description

/code-review <target> --level low does not actually run at low effort. The --level argument is correctly echoed into the "Review target" line of the coordinator's own prompt, but the effort-scaling instructions underneath are hardcoded to the high-effort configuration regardless of what level was requested — so a "low effort" review costs the same as a high-effort one.

Evidence

Ran /code-review HEAD --level low. Reading the coordinator subagent's own transcript, its system/task prompt begins:

Review target: `HEAD --level low`

`high effort → 3+5 angles × 6 candidates → 1-vote verify (recall-biased) → ≤10 findings`

You are reviewing for **recall** at high effort: catch every real bug a careful
reviewer would catch in one sitting. At this level, catching real bugs matters
more than avoiding false positives. Err on the side of surfacing.

## Phase 1 — Find candidates (3 correctness angles + 3 cleanup angles + 1 altitude
angle + 1 conventions angle, up to 6 each)

Run **8 independent finder angles** via the Agent tool...

--level low is correctly captured as the literal review target string, but every actual scaling parameter — "high effort", 8 finder angles, 6 candidates per angle, recall-biased/err-on-the-side-of-surfacing framing — is the same text used in an earlier --level medium run in this same session the day before, word for word. The coordinator then spawned 5 parallel Agent-tool batches covering all 8 angles, same as the higher-effort run.

The tool description for this skill says effort level controls scope ("low/medium: fewer, high-confidence findings; high→max: broader coverage, may include uncertain findings"), so this is a real behavior regression, not a documentation gap — the skill's internal template appears to not be branching on the parsed level at all, or branches incorrectly and falls through to the high-effort text/angle-count in every case.

Impact

This isn't just wrong scope — it's a real cost problem. A "low effort" invocation is supposed to be the cheap option (fewer angles, fewer candidates, no broad-coverage framing), but it silently launches the full 8-angle / 5-subagent-batch high-effort pipeline every time, at full token cost, with no visible indication to the user that the requested level was ignored.

Suggested fix direction

Verify the level-to-scaling-text mapping is actually wired to the parsed --level value (not just to the literal argument string shown in the "Review target" line), and add a quick regression check per level (low/medium/high/xhigh/max) confirming the angle count and framing text actually differ.