fix: load review backend managers only when selected

[atomantic]

Summary

Review-defaults callers imported LM Studio and Ollama download/install dependencies even when they never executed a local review. Load each manager only at the selected backend's existing async boundary, matching MTPLX. The review service's static closure shrinks from 149 to 138 modules.

• Evidence: server/services/codeReview.js:66 owns endpoint dispatch and :359 owns the Ollama capability gate. server/services/agentCliSpawning.test.js:110 and server/services/cleanupAgentWorktree.test.js:14 explicitly document mock accommodations for the former eager manager dependency. Defaults are consumed by agent prompting, task generation, and completion cleanup.
• Ownership/reuse: searched the lib catalog and exports for reviewer configuration, local provider runtime, model probing, and import-graph guards; inspected manager accessors and review importers. Retain manager ownership of live endpoints/capabilities and service ownership of review execution. Reuse call-site dynamic imports and the existing importScoping.test.js:149 guard; no new public surface, catalog, wrapper, or migration.
• Compatibility: public exports, settings cache, endpoint overrides, unknown-capability fallback, and thinking retries retain their contracts. Manager initialization is deferred until use.
• Audit coverage: ranked tracked production source by size and last-120-commit churn, then inspected fan-in and dependency mixing in code review. Checked the supplied open-work snapshot, closed issues by codeReview, merged PRs, and file history. Related feat: offer OpenCode, Kimi and MTPLX as code reviewers #5967/fall back to a local reviewer's served model so an unpinned mtplx review isn't blocked #6081 establish MTPLX behavior; A non-thinking ollama model makes every local reviewer AND the claim-comment gate fail, silently shrinking the claimable queue #5960/Local reviewer sends reasoning_effort to every ollama model, 400-ing on models without thinking support #6050/Local reviewer sends reasoning_effort to models that reject thinking, hard-failing the review and the claim comment gate #6060 cover thinking compatibility, which stays intact. Excluded declarative validation/registry/barrel size and rejected a broader service split because the existing async seam solves this finding without moving exports.

Test plan

• Passed 346 tests across services/codeReview.test.js, services/codeReview.providers.test.js, services/codeReview.backendLoading.test.js, routes/codeReview.test.js, lib/importScoping.test.js, services/agentCliSpawning.test.js, and services/cleanupAgentWorktree.test.js (npm test --prefix server -- --run ...).
• Import guards assert that defaults loading cannot statically reach any of the three managers and that the lazy imports remain present. Existing workflow tests cover successful reviews, failures, pins, capability fallback, and cleanup compatibility.
• git diff --check passed.
• Claude ran in enforced read-only mode for the configured one-round budget. Its response included extra prose (optional no-verdict); the concrete finding was validated and fixed, with a workflow test for manager-load failure at an explicit endpoint.