Design & implement retrieval / context injection

[AlexChesser]

Summary

Design a first-class retrieval primitive for RAG-style context injection. This is NOT IN SPEC and requires spec authoring.

Parent issue: #105 — Missing Modality C

Why

context: shell: and context: mcp: exist, but there's no first-class retrieval primitive — "search these files/this vector store for content relevant to the prompt and inject it." RAG is the dominant accuracy-improvement technique for LLM applications. A context: retrieve: or context: search: type would be high value.

Design Decisions Needed

• Step syntax — new context: retrieve: type? context: search:? Both?
• Search backends — file glob? Full-text search? Vector store? Pluggable?
• Query source — from the step prompt? From a template variable? Explicit query string?
• Result format — raw text injection? Structured with source attribution?
• Result limits — max results? Max tokens? Relevance threshold?
• Whether this should be a built-in or achieved via MCP tools + context: mcp:

Spec Work Required

New spec section needed or extension to existing context types (§6).

Acceptance Criteria

• Spec section authored
• At least one retrieval backend implemented (e.g. file search)
• Retrieved context is injected into step prompt
• Retrieval metadata available in template variables