Static module composition

Optional modules add evidence-backed attributes to the existing graph. They cannot replace nodes, edges, containers, repository identity, or entity labels. The deterministic scanner and core remain usable without optional modules. There is no runtime package loading, remote module discovery, or AI invocation.

Generate-time selection

Enable either or both built-ins in the scanned repository's .topo/config.json:

{
  "schemaVersion": "1.0",
  "repositoryId": "my-repository",
  "modules": ["@topo/module-degree", "@topo/module-cycles"]
}

@topo/module-degree reports incoming and outgoing edge counts, not distinct neighbors. Parallel edges count separately and self-loops count in both directions. @topo/module-cycles reports directed cyclic component membership size: zero outside a cycle, one for a self-loop.

Run topo scan or topo ingest to regenerate. Both use the same composition path before layout, normalized reports, curated views, and the atomic site bundle. Changing config alone does not change an already generated snapshot. Empty modules preserves the core workflow. Unknown and duplicate configured IDs fail; the tool never interprets config strings as import paths.

Each enabled module runs against the same base graph, not another module's outputs. Disabling a built-in removes its generated contributions; rescanning does not retain stale derived values. Source identity and human-authored view definitions remain unchanged. Module changes can change the full graph hash, so an already open editor may need a reload before saving.

Build-time selection

The production site statically includes the two built-in inspector views by default. A distribution can intentionally compile a reduced view catalog:

TOPO_SITE_MODULES="@topo/module-degree" corepack yarn build
TOPO_SITE_MODULES="" corepack yarn build

Unset TOPO_SITE_MODULES includes the whole built-in catalog. An explicitly empty value includes none. Invalid or duplicate IDs fail the build. This selects supported view registrations, not an assertion that every shared library byte has been tree-shaken out.

The data snapshot does not choose executable code. Rebuilding a site changes its supported catalog; regenerating a graph changes its enabled producers. These are independent operations.

Inspector views and version skew

Select a file or external package, then choose Module view in the sidebar. Views expose derived values, method, confidence, and evidence identifiers. They coordinate with the existing map selection without changing its geometry or labels. Metrics describe the full scanned graph, including relationships outside a curated view; collapsed-region values are not fabricated by summing per-node results.

Core schema major-version incompatibility still rejects the entire artifact. Missing module support never authorizes silently interpreting unknown fields. The existing compatibility rule permits older compatible implementations and schemas, but not a newer producer masquerading as an understood version.

Manifest and extension contract

The full manifest is static, first-party code: implementation/schema versions, label, owned attribute fragments, and view registrations. The graph's existing modules array records only {id, version, schemaVersion}; graph schema 1.0 does not gain new fields.

The initial fragment grammar is deliberately bounded: node attributes with nonnegative integer values. It is not arbitrary JSON Schema execution in the browser. Catalog validation rejects conflicting attribute/view ownership, unsupported fragments, invalid versions, and module dependencies. Modules depend on the core only; adding another module requires a source change, a declared manifest, validation, and a new site build.

See the pure core API, generation integration, and schema compatibility contract.

Workload acceptance

Implementation df8bf5157623ba3c07d4d80165bc62e933e34044 completed all 24 frozen Mermaid/VSCode observations above the agreed 30 whole-workload FPS floor. The source-bound manifest retains every capture, raw interval, verified interaction effect, module readiness check, phase observation, and submitted buffer range.

Largest expanded VSCode workload:

The minimum across all observations was 46.365 FPS. Long frames remain in the score, including a 597.225 ms frame. All interaction effects were verified and no browser errors occurred. Rendering remains uncapped.

These are acceptance observations, not isolated speedup evidence. Curated views exclude external/synthetic nodes. Headed near-vsync intervals were about 6.94 ms versus 16.67 ms headless; cross-mode FPS is not a renderer comparison. Optional analysis also has a real generation cost: composing both modules added about 3.94-4.38 seconds to the expanded-source VSCode fixture's preparation. That cost is reported separately, not hidden in the frame score.

These captures belong to the retired WebGL benchmark archive. The original harness and external frozen graph inputs are no longer part of the maintained tree, so the measurements are preserved evidence rather than a current reproduction recipe.