feat: adopt skills open standard (agentskills.io)

Description

Epic for the open-standard / multi-harness effort. https://agentskills.io/home

Strategy: publish skills to the open agent-skills standard (.agents/skills/ convergence + skills.sh registry) instead of maintaining per-harness plugin manifests. Tiers:

• Tier 1 — Claude Code: full plugin (skills + hooks + agents + MCP). Primary, always supported.
• Tier 2 (wave 1) — Codex + Gemini CLI + OpenCode: all three read .agents/skills/, support MCP, and support subagents. Codex and Gemini ship Claude-compatible hook systems (JSON-over-stdin, same PreToolUse/PostToolUse/UserPromptSubmit events), so the enforcement layer ports with adaptation. OpenCode uses a JS/TS plugin API instead, so its hooks are a rewrite. CI-tested.
• Experimental: every other harness that reads .agents/skills/ (Cursor, Antigravity, Copilot, ...) — skills work for free, no dedicated CI or guarantees.

Linked under this epic

• #50 release in skills.sh
• #49 support AGENTS standard
• #46 Codex support (tier 2) — skills + MCP (config.toml) + hooks (hooks.json / [hooks], feature-flagged, no Windows)
• #125 Gemini CLI support (tier 2) — skills + MCP + hooks (via Gemini's Claude-hook migration tooling)
• #127 OpenCode support (tier 2) — skills + MCP + subagent; hooks rewritten as OpenCode plugin
• #126 Phase 0 portability prep — guard MCP references, flatten namespacing + cross-plugin collision audit, $ARGUMENTS / ${CLAUDE_SKILL_DIR} fallbacks, hook-adaptation for Codex/Gemini (blocks the rest)

Superseded / downgraded to experimental: #47 Antigravity, #48 Cursor (per-harness manifests are no longer the approach).

Requirements

• All CI/CD checks are passing.
• There is no drop in the test coverage percentage.

Additional Context

No response