Refactor: split monolithic learning_layers/harness.py and resolve core->layers import inversion

[bordeauxred]

Why

`clawloop/learning_layers/harness.py` is 1139 lines and mixes five conceptually distinct things in one file:

• `ToolConfig`
• `PlaybookEntry` / `Playbook` (ACE-style memory)
• `PromptCandidate` / `ParetoFront` (GEPA)
• `Insight` (reflector output validation)
• `Harness` itself (Layer-protocol orchestration)

A reader trying to grok the layering can't see the boundaries.

There's also an architectural smell: `clawloop/core/curator.py:25` imports from `clawloop/learning_layers/harness.py`. Core depending on a layer is the wrong direction; `Playbook`/`Insight` should live in core (they're domain types, not layer mechanics).

What

• Move `Insight`, `Playbook`, `PlaybookEntry` from `learning_layers/harness.py` to a new `clawloop/core/playbook.py`. Update `curator.py`'s import.
• Split GEPA pieces: extract `PromptCandidate`, `ParetoFront`, `_dominates` into `clawloop/core/pareto.py` (pure data + algorithm) or `learning_layers/harness_pareto.py` (if it stays layer-coupled).
• Extract `ToolConfig` into a small module (`clawloop/core/tool_config.py`) — it's referenced by multiple layers.
• Leave `harness.py` as a Layer-protocol orchestrator only. Target ≤350 lines.
• Update all imports across `clawloop/`, `tests/`, and `examples/`.
• All existing tests pass with no behavioral change.

Acceptance

• `grep -r "from clawloop.learning_layers.harness import" clawloop/core/` returns nothing
• `learning_layers/harness.py` is ≤400 lines
• `pytest tests/ -x -q` green

Effort

~2-3 hours. Mechanical refactor; risk is import cycles.

Source

Code audit and Codex review, 2026-04-18. Both auditors independently flagged this as the top S1 refactor.