fix(review): ground findings in current code and full PR context

Problem

The shared code-review skill and PR-review plugin produce avoidable false positives and miss important context. The skill asks for arbitrary complexity, naming, and improvement comments even though repository guides reject that noise. Optional file-review subagents emit minor and nit findings. The plugin preloads reviews and inline threads but omits top-level PR discussion and linked issue acceptance criteria.

This matches observed review failures: approved PRs later received concrete human objections already present in top-level comments or issue criteria, while bad findings concentrated in optional refactors, style, documentation demands, and speculative tests.

Desired Behavior

Both Canvas-driven reviews and the reusable PR-review plugin should ground decisions in the current head, linked acceptance criteria, existing discussion, and the repository's current implementation, while reporting only material failures.

Acceptance Criteria

• Define a material finding as a concrete current-head failure with traceable evidence.
• Require the reviewer to understand the existing or adjacent implementation before judging the patch.
• Include linked issue bodies and top-level PR comments in review context, ahead of historical/resolved context.
• Require re-verification of previous human concerns and current-head CI.
• Remove minor/nit output from file-review subagents and require coordinator re-verification.
• Preserve security, compatibility, dependency freshness, evidence, and risk assessment checks.

Locate the shared code-review skill, the PR-review plugin, and the optional file-review subagents, then trace how review context and findings are produced. Compare current-head, linked-issue, top-level-comment, and CI handling against the acceptance criteria. Done means material findings are evidence-based, minor output is removed, context is complete, and previous concerns are re-verified.