continue: translate canonical Agents into .continue/agents/*.yaml config format

[scottmeyer]

Status: deferred from v0.8.0 plugin upgrade.
Plugin: continue (plugins/continue_plugin.go)
Capability cell: AGENTS, currently ----

Context

Continue's "custom agents" exist but in a config-shape, not a file-shape:

• Location: ~/.continue/agents/*.yaml (user-global) or .continue/agents/*.yaml (project)
• Format: full YAML config composing models + rules + tools + system prompt
• Reference: https://docs.continue.dev/reference

This is a non-trivial mismatch with the canonical model where model.Agent is markdown + frontmatter (mirrors Claude/Cursor/Gemini/Copilot). A real translator needs to:

1. Synthesize a YAML config from the canonical Agent's markdown body, description, and frontmatter
2. Decide how to bind models / rules / tools (defaults from project + frontmatter overrides)
3. Handle the inverse direction in the importer

v0.8.0 left this at ---- because the translator is mid-sized and benefits from a thoughtful design pass rather than a fast port.

What to implement

1. Design the canonical → Continue agent config mapping:
   • Agent.Description → Continue's description
   • Agent.Document.Body → Continue's prompt or instructions
   • Frontmatter fields (model, temperature, etc.) → Continue's models[] + defaultModel
   • Agent.Tools (wildcards or explicit) → Continue's tools[] config
2. Emit .continue/agents/<n>.yaml in plugins/continue_plugin.go's Plan().
3. Update Capabilities().Agents from SupportUnsupported → SupportDegraded (the translation loses some structure) or SupportNative if a clean mapping is found.
4. Importer parity: extend internal/importer/continue.go to read .continue/agents/*.yaml back into model.Agent. The reverse mapping should round-trip — at minimum body + description + name.

Acceptance criteria

• prism capabilities shows continue AGENTS = degr. or native.
• A canonical agent projects to a working Continue agent config.
• Round-trip test: agent fixture survives init → compile → check.

Open design questions

• How to handle Continue-specific tooling (rules + models + MCP composition) when the canonical agent doesn't carry those?
• Should the YAML be lockfile-tracked, or treated as a user-editable file (like .continue/permissions.yaml)?