refactor: reduce internal barrel imports

[jayscambler]

Summary

• split direct implementation modules out from broad barrels for the LLM judge and codegen registry
• switch several internal runtime/control-plane imports to the new direct modules
• keep public barrel exports stable while reducing internal coupling

TDD / DRY framing

This is the next stacked architectural PR after #650.

Test-first work:

• added direct-import regression tests first
• extracted implementation modules
• rewired internal consumers to import directly
• preserved external/public re-export surfaces through existing barrel files

Changes

New direct modules

• ts/src/judge/llm-judge.ts
• ts/src/scenarios/codegen/registry.ts

Barrel files kept as public compatibility surfaces

• ts/src/judge/index.ts
• ts/src/scenarios/codegen/index.ts

Internal imports moved off barrels in runtime paths

• ts/src/execution/task-runner.ts
• ts/src/mcp/server.ts
• ts/src/scenarios/agent-task-factory.ts
• ts/src/investigation/engine.ts
• ts/src/simulation/engine.ts
• ts/src/knowledge/solver.ts

New tests

• ts/tests/internal-module-imports.test.ts

Validation

• cd ts && npx vitest run tests/internal-module-imports.test.ts tests/judge.test.ts tests/codegen.test.ts tests/task-runner.test.ts tests/generation-loop.test.ts
• cd ts && npm run lint

Follow-up

After this PR, the major remaining architectural recommendation is the state-transition / state-machine work for GenerationRunner and mission workflows.