feat(framework): support scoped sub-agents across five hosts

[alexsoyes]

What & why

Add optional, self-contained agents installed by the CLI with an explicit provider and IDE.

Main changes

• Opt-in: install or remove four native roles in the current project with aidd agents. Plugins install none by default.
• Native compatibility: Codex inherits its parent model for Claude shorthand selectors; Copilot agents receive repository instructions, with explicit opt-out preserved.
• Ownership: concurrent mutations serialize per project, including cleanup. Failed writes restore the previous native files; dangling personal links are rejected before writes; removal targets only the selected IDE.
• Roles and generation: explorer, executor, validator and reviewer each define one behavior and output in one file per host, without required skills or companion files. The executor favors simple, minimal code and documentation; the reviewer recommends removing obsolete or redundant material.
• Beta guide: README opt-in entry and subagent guide with roles, local build, native trial, removal by IDE, beta limits and a Markdown feedback template.
• Plan: explore independent codebase questions in parallel when useful.
  • Prefer independent phases with separate write ownership.
• Dependencies: record prerequisite outputs, overlapping writes and shared mutable checks in Depends on.
• Implementation: accept a whole plan or one assigned phase; the caller integrates and finalizes once. Batch checks the combined result.
• Models: one CLI JSON lists every role and assigns standard or light provider profiles for OpenAI, Anthropic and Mistral. Main and review share the configured model and effort; main-session settings remain unchanged. Mistral targets OpenCode V2.
• Workflows and context: delegate by responsibility using available host agents or a scoped general-purpose agent, without custom agent names or installation dependencies.
• Reload: make reload syncs all plugins across five hosts and generic global delegation guidance. Standalone agents stay untouched; OpenCode preserves its actual Claude instruction fallback despite Claude path overrides.

Verification

• Automated: 6,815 CLI tests pass, one skipped; 502 script tests pass. Types, lint, architecture, unused-code, links and hooks pass. Bundle: 710.4 KB against a 715 KB budget.
• Native files: 25 built-CLI cases cover eight provider/IDE pairs, per-role model and effort, repeat installation, provider changes, removal, rejected selections, personal-file conflicts including dangling links, cleanup and a distribution without agents.
• Smoke: 140 checks pass, one remote-fetch check skipped; all 34 leaf commands exercised.
• Regression guards: Targeted mutations disabling ownership, rollback, serialization, instruction inheritance and model filtering fail their named tests. Dangling native and registry links are preserved before writes; default filesystem behavior remains unchanged.
• Reload regressions: reproduced personal-agent deletion and incorrect OpenCode fallback before correction; all 47 reload cases pass.
• Evidence and limits: agent guide, versioned verification. Headers do not prove account access or effective runtime model selection. Earlier runtime and cost probes remain historical.

Gather, wireframe and generic skill generation remain unchanged against next. Generic skill generation changes stay in #1028. Parallel execution remains optional.

Refs #983.

I certify

• I DO CERTIFY I READ EACH LINE OF THE PULL REQUEST BECAUSE I AM A SOFTWARE ENGINEER, NOT A AI PUPPY.