feat: evolutionary prompt optimization — Rainbow Teaming QD loop for role system prompts

[justrach]

Problem

Role system prompts in src/runtime/roles.zig are currently static and hand-tuned. There's no systematic way to evaluate whether a prompt is optimal for its role, or to evolve prompts over time based on actual task performance.

Proposal

Implement an evolutionary prompt optimization loop inspired by quality-diversity (QD) search methods, specifically the Rainbow Teaming approach (arXiv:2402.16822).

Core idea

Cast prompt optimization as a quality-diversity problem:

• Quality: task success rate, code correctness, fix accuracy
• Diversity: coverage across different task types, codebase patterns, failure modes

Use open-ended search to generate prompt variants that are both effective and diverse, maintaining a MAP-Elites style archive of best prompts per niche.

Implementation sketch

1. Prompt genome: each role's system prompt is a "genome" that can be mutated
2. Fitness function: run the role on a benchmark task suite, measure success metrics
3. Diversity dimensions: task category (bug fix, review, search), codebase size, language features used
4. Selection: MAP-Elites archive keeps the best prompt per (quality, diversity) cell
5. Mutation operators: LLM-guided rewriting — "make this prompt better at X while keeping Y"
6. Evolution loop: generate variants → evaluate on benchmarks → archive best → repeat

Integration with grid

The evolved prompts feed back into the grid system:

• grid.zig currently maps role → model tier
• Extend to map role → (model tier, prompt variant ID)
• Store winning prompts in .devswarm/evolved_prompts/ per project
• Fall back to built-in defaults when no evolved prompts exist

Connects to

• feat: evolutionary grid tuning — DGM outer loop tunes role→model per codebase #274 (evolutionary grid tuning — DGM outer loop)
• This extends feat: evolutionary grid tuning — DGM outer loop tunes role→model per codebase #274 from model selection to prompt content optimization
• feat: evolutionary prompt optimization — Rainbow Teaming QD loop for role system prompts #353 (eval framework for prompt selection)

Why P0

The swarm now has 12 roles with structured prompts (#352). The next leverage point is making those prompts self-improving rather than hand-tuned. This is the foundation for the entire adaptive agent system.

[justrach]

Scope update per #355

HyperAgents paper shows that prompt-only evolution (what this issue currently proposes) is equivalent to the DGM-H w/o self-improve baseline — the weaker variant. The meta-level (orchestrator decomposition, synthesis strategy) should also be evolvable.

However, for production cost reasons (#355 proposes periodic tuning), prompt evolution is still the right first step. Just need to architect it so we can extend the genome later.

Depends on #355 (architecture decision) and #354 (eval framework).