refactor(internal): audit package cohesion across internal/

[grantcarthew]

Walk each package under internal/ and check whether the files inside still match the package's stated purpose, or whether responsibilities have drifted in as the codebase has grown.

Packages to audit:

• cache
• cli
• config
• cue
• detection
• doctor
• modules
• orchestration
• registry
• shell
• temp
• tui

For each package, ask:

1. Does every file in the package contribute to the role the package name implies, or does some content belong elsewhere?
2. Are any single files doing two jobs (e.g. a "primary type + a grab-bag of utility functions other packages reach into")? If so, would a clean split improve discoverability?
3. Are there parallel type hierarchies across packages that describe the same underlying data — and if so, is the split still earning its keep?
4. Is the dependency direction acyclic and one-way (UI/CLI → runtime → utilities), with no back-edges from lower layers to higher ones?
5. Are public symbols on the package narrow and intentional, or accidental (an internal helper that became Exported because one cross-package caller needed it)?

Out of scope:

• Behaviour changes. This audit is structural only.
• The duplicate-models-walk hoist (refactor(cue): hoist agent models walk into shared helper #2) — that's the right fix in its own right and shouldn't wait on the audit.

Deliverable:

• A short note per package describing what's in scope vs out of scope, what would move, and what stays. Treat it as an inventory pass before any restructuring PR — the moves themselves can land incrementally and don't need to be one giant change.

Surfaced by a pre-commit review conversation (.start/reviews/2026-05-16-pre-commit-01.md).