Deliver review summaries through the native MCP tool

[gontzess]

Why

Sending complete review summaries inside shell heredocs can exceed Claude Code's command-analysis limit or trigger shell-safety approval, leaving a shortened report without its AI-agent appendix. Agent mode also stopped automatically supplying the PR description, making author intent depend on a separate model-initiated query.

Changes

• Reuse the existing upstream mcp__github_comment__update_claude_comment tool; no custom MCP server or new dependency.
• Before the model runs, validate repository/PR/checkout/live-head identity, create a fresh provisional working comment and bind its ID through CLAUDE_COMMENT_ID.
• Send provisional and final Markdown as tool data, retaining upstream sanitization and secret redaction. GitHub's comment-size limit still applies; the shell's command-length limit does not.
• Preserve the complete author-written PR description as pr_body from the existing PR metadata request. Read title/description as untrusted claims to verify against the diff, never as review instructions or verdict policy. No extra API call or permission is needed.
• Preserve CI-owned completed reports, exact-head verdicts, finality checks and completion-before-supersession. Ambiguous working-comment creation fails closed rather than reusing another run's slot or blindly retrying.

No model, effort, runtime pin, background-task policy or shell permission expansion relative to #132. Existing broad gh api capability remains; this is not credential isolation.

Latest validation

• 138 reviewer tests passed locally. Four preparation tests caught 13 targeted compiling mutations.
• Before/after context smoke reproduced the old description omission, then verified complete 23,400-character Markdown, empty/null descriptions, preservation through preparation and unchanged API-request counts.
• A real full 89-finding workflow replay passed on this exact PR head. The entire 1,354-character PR description matched GitHub and appeared in the successful context-read output, independent of an additional description/CI-status query.
• Native MCP delivered both summaries. The 14,416-character final body, including the AI-agent appendix, was retained intact in the archived working comment. Foreground subagents completed, then CI published the completed report and exact-head CHANGES_REQUESTED verdict before superseding prior output. No shell summary writes or weakened publication guards.
• Scoped source/security review found no blocker. These checks validate input delivery, transport and publication—not the correctness of every model finding or immunity to prompt injection.