Add configurable models for memory and skill background reviews

[gregoryschlepper]

Adds optional memory.review.* and skills.review.* overrides so background reviews can use a different model/provider than the main agent. Default behavior remains unchanged.

[gregoryschlepper]

This PR addresses the first half of the review-cost problem.

What is already concretely improved:

• background review can be routed to a cheaper review model instead of always inheriting the main model

What is not yet instrumented:

• exact before/after token usage
• exact latency deltas
• exact 429 reduction

Those likely improve as a consequence, but I am intentionally not overstating that without dedicated review-level telemetry.

I split the optimization work so review-model selection and review-size reduction can be discussed independently.

[gregoryschlepper]

Follow-up note: I also did a static verification pass on the current live branch version of this change.

What is statically confirmed:

• memory.review / skills.review are loaded inside _spawn_background_review
• the effective review runtime values are derived into eff_model, eff_provider, eff_base, and eff_key
• the review agent is instantiated with AIAgent(model=eff_model, provider=eff_provider, base_url=eff_base, api_key=eff_key, ...)
• there is no separate dedicated override later in the review path that resets the review agent back to self.model

So code-wise, the override path is wired through as intended.

What is still not separately instrumented:

• dedicated runtime telemetry for review-only model usage / token accounting

[teknium1]

Thanks for the focused work on lowering background-review cost. This automated hermes-sweeper review found that current main already provides the requested behavior through the shared auxiliary-task configuration surface.

• 87c4a5ebb8a9f8122197a908288cc0abc7cef6b0 (feat(background-review): aux-model selector for the self-improvement review, feat(background-review): aux-model routing + context digest + adaptive cadence to cut self-improvement cost #49252) added auxiliary.background_review as the canonical review-routing slot.
• hermes_cli/config.py:1712 defines provider, model, base URL, API key, and timeout settings; unset/auto retains the main-model behavior.
• agent/background_review.py:46 resolves the configured review runtime, and agent/background_review.py:663 constructs the review agent with the resolved model/provider/credentials.
• tests/run_agent/test_background_review_cost_controls.py:1 covers the configured-routing and fallback behavior.

This supersedes the memory.review / skills.review implementation proposed here, so closing as implemented on main.