feat: evolutionary grid tuning — DGM outer loop tunes role→model per codebase

[justrach]

Parent: #257 (Phase 2)

What

The evolver (#148) meets the model grid. The base grid is a static mapping, but which model actually performs best for each role in your codebase is an empirical question.

The DGM (Darwinian Grid Manager) outer loop:

1. Logs turn quality signals (fitness scores, review pass/fail, diff quality)
2. After N runs, analyzes which model performed best per role
3. Writes optimized .codedb/model_grid.toml with rationale comments
4. The grid improves itself over time

Codegraff Pro feature.

Depends on

• feat: grid override file — .codedb/model_grid.toml for per-role overrides #269 (grid override file)
• [Epic] Evolutionary Self-Improving Code Agent (run_evolver + evolve_agent) #148 (evolutionary agent epic)
• feat: model grid — base Role→Model static mapping #265 (model grid)

[justrach]

Blocked by: #265.

[justrach]

Update: prompt evolution pipeline taking shape

New related issues that extend this work:

• feat: evolutionary prompt optimization — Rainbow Teaming QD loop for role system prompts #353 (P0) — Evolutionary prompt optimization using Rainbow Teaming QD search (arXiv:2402.16822). Extends this issue from model selection to prompt content optimization.
• research: multi-agent eval framework — what metrics select the best role prompts? #354 (P0) — Multi-agent eval framework to define fitness functions for prompt selection. Prerequisite for both this issue and feat: evolutionary prompt optimization — Rainbow Teaming QD loop for role system prompts #353.
• Upgrade swarm: role-aware dispatch, specialist roles & structured system prompts #352 (PR) — Adds 12 structured role prompts + specialist roles to the grid. This is the baseline the evolution loop will optimize from.

Proposed execution order: #354 (define evals) → #353 (evolution loop) → #274 (full DGM outer loop with model + prompt co-optimization).

[justrach]

Blocked by #355

New research (HyperAgents, arXiv:2603.19461) changes the scope of this issue. Before we implement evolutionary grid tuning, we need to decide:

1. Do we evolve only task agents (DGM-style) or the full stack including meta-agents (DGM-H-style)?
2. What's the tuning cadence for production? (continuous vs weekly/monthly)
3. What's in the genome? (prompts only vs prompts + grid + decomposition strategies)

See #355 for the full analysis. This issue is blocked until that architectural decision is made.