tooling: centralize machine-readable architecture ownership declarations

[thymikee]

Outcome

Create one machine-readable architecture-ownership declaration consumed by layering and depgraph tooling. Remove duplicated logical-module and snapshot-policy roots without changing production imports or runtime behavior.

Why

The authority-overlay plan currently names vocabulary, capability, live-state-shape, live-state-authority, and executable-policy, but it does not declare the exact roots/exports that make vocabulary and capability non-heuristic. This prerequisite establishes that evidence source before the graph is extended.

Scope

• Add scripts/layering/architecture-ownership.ts.
• Move or re-export LOGICAL_MODULE_POLICIES roots from scripts/layering/daemon-modularity.ts.
• Declare src/snapshot/ as the executable-policy owner with its existing forbidden daemon target.
• Declare exact initial vocabulary roots/subpaths and capability roots/exports used by the follow-up graph issue. Every entry must point to an existing enforced owner or public capability; no filename guessing.
• Update existing layering tests to consume the shared declarations.
• Delete superseded duplicate tables/constants.

Out of scope: production imports, depgraph JSON changes, numerical thresholds, moving snapshot/replay/handler code.

Size constraint

This is a consolidation. Aim for net production/tooling LOC at or below zero after deleting duplicate declarations. Any net growth must be itemized and justified in the PR.

Proof

1. Plant a snapshot-policy import of src/daemon/types.ts; pnpm check:layering must fail naming the existing boundary.
2. Plant one invalid logical-module edge; it must fail through the shared declaration.
3. Add exact unit coverage showing every declared classification root resolves and similarly named undeclared files do not classify.
4. Run focused layering tests and pnpm check:affected --run.

Done criteria

• One declaration owns logical roots, snapshot-policy ownership, and the initial exact vocabulary/capability evidence.
• Existing gates retain their semantics and positive controls.
• Duplicate declarations are deleted.
• No production behavior or import changes.
• Planted-red evidence, touched-file count, and net LOC are reported in the PR.

Stop conditions

• A proposed capability has no existing production owner or real consumer.
• Sharing declarations would require a second parser or loosen an existing gate.
• The change grows into the depgraph payload work.