fix(code-review): strengthen step 1 gating agent reliability

[kpatel513]

Summary

Step 1 used a Haiku agent with no defined criteria to decide whether to skip a PR as "trivial". A wrong skip silently drops the entire review with no output. Two issues fixed:

1. Wrong model: Haiku is the weakest model for a binary decision that gates the entire workflow. Upgraded to Sonnet.
2. Vague skip criteria: "trivial change that is obviously correct" is undefined — Haiku would interpret this inconsistently. Replaced with explicit criteria.
3. Fragile "already reviewed" check: "comments left by claude" would match any comment mentioning the word "claude". Changed to look for the ## Code review header that the plugin actually posts.

Changes

plugins/code-review/commands/code-review.md

• Step 1: Haiku → Sonnet
• Step 1: Replace vague "trivial" description with explicit skip criteria (lock files, changelogs, generated code, whitespace-only, automated dependency bumps)
• Step 1: Replace "comments left by claude" with check for ## Code review header

Before / After

Before:

Launch a haiku agent to check if...

• The pull request does not need code review (e.g. automated PR, trivial change that is obviously correct)
• Claude has already commented on this PR (check for comments left by claude)

After:

Launch a sonnet agent to check if...

• The pull request is trivial... A PR is trivial only if it exclusively contains: auto-generated file changes, whitespace/formatting with no logic changes, or automated dependency bumps. If the PR contains any non-trivial changes alongside these, it is not trivial.
• Claude has already reviewed this PR (check for a comment containing the ## Code review header)

Test plan

• Run on a Dependabot PR — should be skipped
• Run on a PR with a Dependabot bump + one logic change — should NOT be skipped
• Run on a PR already reviewed — should be skipped (check via ## Code review header, not author name)
• Run on a substantive PR — should proceed to review