feat: add code review pipeline to rhdh-pr-review skill (v0.3.1)

[schultzp2020]

Summary

Consolidates the personal pr-review-github skill into rhdh-pr-review with a layered, forge-pluggable architecture.

Architecture

fetch-github.md ──→ review-code.md ──→ post-to-github.md
       │
       └──────────→ review-operator-pr.md

# Future:
fetch-gitlab.md ──→ review-code.md ──→ post-to-gitlab.md

Three layers, each with a single job:

• Fetch (forge-specific): collects PR metadata, diff, linked issues, existing comments, CI status → context artifact
• Analyze (agnostic): runs review perspectives, verifies findings → findings artifact
• Post (forge-specific): maps findings to inline review comments → posts to forge

The analyze layer never calls forge-specific CLIs (except reading file contents at HEAD). Adding GitLab support later requires only fetch-gitlab.md + post-to-gitlab.md — no changes to the analysis workflow.

New files

File	Purpose
scripts/fetch_pr_context.py	Stdlib-only script: PR URL/number → JSON context artifact
workflows/fetch-github.md	Documents script usage and context artifact contract
workflows/review-code.md	Platform-agnostic analysis with soft perspective guidance
workflows/post-to-github.md	Findings artifact → GitHub inline review via gh api
references/review-perspectives.md	Example review perspectives and signal hints

Modified files

File	Change
SKILL.md	Rewritten as router with 4 routes + artifact contracts
workflows/review-operator-pr.md	Phase 1 now consumes shared context artifact
skills/rhdh/SKILL.md	Updated routing triggers + skills index
README.md	Updated skill description

Design decisions

• Review perspectives are soft guidance, not a fixed roster — the agent creates perspectives based on what matters for each PR
• Findings are severity-ranked: medium+ findings get inline comments automatically, nits/low items are listed as one-line bullets for the user to cherry-pick
• No double-confirmation: if the user approved findings in review-code.md, post-to-github.md does a brief confirm only
• Follow-up reviews: top-level comment can be skipped when a prior review exists, but not a hard rule

Review process

• Drafted all files
• Ran skill-reviewer subagent — fixed all 4 medium issues, 4 low issues, applied 2 suggestions
• Passed skill-maker Phase 5 audit checklist
• All pre-commit hooks pass (ruff, markdownlint, pytest — 283 tests)