feat(ci): mechanical module-boundary enforcement (eslint-plugin-boundaries/dependency-cruiser + madge) (CA-03)

[krokoko]

This is a finding from https://github.com/krokoko/cairn (action item CA-03).

Component

Tooling / CI

Describe the feature

Add mechanical module-boundary enforcement that encodes the AGENTS.md routing table as code, using eslint-plugin-boundaries or dependency-cruiser, plus madge --circular for circular-dependency detection. Start the boundary rules in warn mode, snapshot the current baseline, then flip to error once the baseline is clean.

Boundaries to enforce (from AGENTS.md):

• Handler layer (cdk/src/handlers/) can't import stack/construct internals (cdk/src/stacks/, cdk/src/constructs/).
• CLI (cli/src/) can't import CDK internals.
• No circular dependencies anywhere.

Use case

The AGENTS.md layering rules are prose, not mechanically enforced (AI007 / AI003). Nothing stops an agent — or a human — from introducing an import that violates the intended architecture, and circular dependencies go undetected. A boundary linter turns the routing table into a binding fitness function that catches architectural drift at PR time instead of in review.

Proposed solution

1. Add eslint-plugin-boundaries (or dependency-cruiser) to cdk / cli and encode the AGENTS.md routing table as boundary rules.
2. Add madge --circular as a fast check (pre-commit + CI).
3. Land the rules in warn mode; record the current violation baseline.
4. Once the baseline is clean, flip to error and add to the required build check.

Acceptance criteria

• Boundary rules encode the AGENTS.md routing table (handler↛stack/construct internals; CLI↛CDK internals).
• madge --circular runs in pre-commit and CI; zero circular deps in the clean baseline.
• Rules start in warn; baseline violations are documented.
• A follow-up flips clean rules to error in the required build check.
• Remediation text on failures is agent-targeted (what to fix, not just what failed).

Other information

Source reports: readiness-roadmap.md, verification-strategy.md (§"Architecture Fitness Functions", maturity 0→3), ai-smells-gates-report.md (AI003/AI007). Effort: S–M. Calibrate permissively — agents route around hard blockers, so introduce the error line only after the code is behind it. Per ADR-003 this issue needs the approved label before work begins.