fix: align managed capabilities across agent harnesses

[kapelame]

What & why

Agent capability support was inconsistent between the create flow, server validation, renderer, and daemon adapters. Codex and OpenCode rejected portable Skill archives even though their runtimes can load managed Skill roots, while unsupported capabilities could surface as the misleading capability_credential_missing runtime error.

How

• validate capability compatibility when creating, editing, or enabling an Agent capability
• classify unsupported capabilities as capability_unsupported instead of a credential error
• materialize managed Skills under stable ~/.parsar/runtime/<engine>/.../skills state
• register managed Skill roots through Codex skills/extraRoots/set and OpenCode skills.paths
• preserve Pi Skill delivery through repeated --skill <dir> flags; Pi MCP remains unsupported
• show the selected Agent engine in the capability step

Harness	Skill	MCP
Claude Code	Supported	Supported
Codex	Supported by this change	Supported
OpenCode	Supported by this change	Supported
Pi	Already supported	Unsupported

Verification

• make check
• Relevant E2E target: no dedicated cross-harness capability E2E target exists
• Adapter and connector regression tests cover Skill materialization, registration, compatibility validation, and unsupported error classification

Notes for reviewers

This is stacked on #249 so the diff contains only the three harness compatibility commits. It does not include the closed MiniMax authentication PR or capability editor changes.

GitHub-hosted jobs currently fail before executing any steps because the organization account is locked due to a billing issue; the full local make check gate passed.