[azsdk-cli code review] Author the skill structure and review contract

## Parent - #16764 ## Goal Author the first review-focused skill for `tools/azsdk-cli/**` with a small, maintainable instruction surface and an explicit finding contract. ## Deliverables - [ ] Add `.github/skills/code-review/SKILL.md` with valid frontmatter and explicit review-task triggers. - [ ] Scope azsdk-cli-specific review behavior to diffs under `tools/azsdk-cli/**`. - [ ] Add progressive references for the conditional architecture/risk map and finding contract. - [ ] Define which findings belong to the skill versus compiler, MCP001-MCP007 analyzers, formatting, and CI. - [ ] Require each finding to identify an exact changed line, a concrete failure path, and user impact. - [ ] Add trigger and anti-trigger evals, including unrelated repository changes and analyzer-only controls. - [ ] Reconcile duplicated guidance in the existing reviewer agent so review rules have a clear source of truth. - [ ] Run Vally lint, reference checks, token checks, formatting, and the targeted trigger evals. ## Content principles - Keep `SKILL.md` procedural and concise; load detailed guidance only when relevant. - Prefer links to authoritative repository context over copied architecture prose. - Do not report pre-existing, style-only, theoretical, or analyzer-owned findings. - Review behavioral and cross-file risks: CLI/MCP parity, DI/service lifetimes, cancellation, bounded external calls, response contracts, mock drift, unsupported-language paths, and failure-path tests. - Every retained rule must identify its source/owner and later receive a positive or negative benchmark fixture. ## Acceptance criteria - Skill passes repository validation and token budgets. - It activates for azsdk-cli PR review prompts and stays inactive for unrelated changes. - Example output contains only actionable, line-level findings using the finding contract. - A maintainer can update architecture guidance without editing the same rule in multiple places.