feat(skills): add RTK_DISABLED=1 opt-out skill and install via rtk init

[bb1]

Summary

• Add an agent-agnostic skill (skills/raw-command-output/SKILL.md) that teaches LLM agents to prefix commands with RTK_DISABLED=1 when they need raw, uncompressed output (JSON/XML/CSV for data analysis)
• Install the skill to ~/.agents/skills/ during rtk init --opencode (and combined Claude+OpenCode modes)
• Show skill path in rtk status output; remove on uninstall

Motivation

When rtk compresses command output for token savings, agents sometimes cannot parse the result (e.g., JSON data for analysis). This causes infinite retry loops where the agent keeps re-issuing the same command with slight variations.

The RTK_DISABLED=1 env prefix already exists in the Rust binary (#345) but agents don't know about it. This skill teaches them to use it proactively.

Related Issues

• Closes Feature request: per-invocation escape hatch (flag/env var) so an agent can self-disable rtk rewriting when stuck #1777 — per-invocation escape hatch so agents can self-disable rtk when stuck
• Relates to Output filtering causes AI agents (Claude Code) to burn extra tokens on retry loops #831 — output filtering causes agents to burn tokens on retry loops
• Relates to rtk git branch --show-current loop in opencode #732 — rtk git branch --show-current loop in OpenCode
• Relates to enhancement: add RTK_DISABLED=1 environment variable to bypass all hooks #1153 — RTK_DISABLED=1 env var to bypass hooks (already implemented in binary; this PR makes agents aware of it)

Changes

File	Change
skills/raw-command-output/SKILL.md	New: agent-agnostic skill file
src/hooks/constants.rs	Add AGENTS_DIR, SKILLS_SUBDIR, RTK_SKILL_DIR, RTK_SKILL_FILE constants
src/hooks/init.rs	Add rtk_skill_path(), ensure_rtk_skill_installed(), update install/remove/status flows

Testing

• cargo check passes (verified via Docker rust:latest)
• Skill installs to ~/.agents/skills/raw-command-output/SKILL.md
• Existing tests unaffected (skill install is home-relative, orthogonal to tempdir-based plugin tests)

[CLAassistant]

All committers have signed the CLA.

[bb1]

@pszymkowiak — would appreciate your review on this when you get a chance.

Quick context: this adds an agent-agnostic SKILL.md that teaches agents to use RTK_DISABLED=1 when they need raw command output (e.g. after retries or when parsing JSON/XML). rtk init installs it to ~/.agents/skills/ (OpenCode) and ~/.claude/skills/ (Claude Code).

One design question: currently only OpenCode and Claude Code support the skills/<name>/SKILL.md format for global skill discovery. Other agents use different conventions (Cursor .cursor/rules/*.mdc, Cline .clinerules/, Windsurf .windsurfrules, Codex AGENTS.md, etc.). Is it enough to support just Claude Code and OpenCode for now, or should we also convert the skill into agent-specific rule formats for the others?